diff --git a/results/env.r1.1465528.txt b/results/env.r1.1465528.txt new file mode 100644 index 0000000..5747bb9 --- /dev/null +++ b/results/env.r1.1465528.txt @@ -0,0 +1,353 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=64801 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier08191 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier08191 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696894702 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=c21tbCIsCndWID0gIl4wMDAwMDAwMC4wMDAwMDAwMDIuMDAwMDAwMDAyLip6ZmluYWwiLAp9LApbImNyYXktbGlic2NpIl0gPSB7CmZuID0gIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NvcmUvY3JheS1saWJzY2kvMjIuMTIuMS4xLmx1YSIsCmZ1bGxOYW1lID0gImNyYXktbGlic2NpLzIyLjEyLjEuMSIsCmxvYWRPcmRlciA9IDEyLApwcm9wVCA9IHt9LApzdGFja0RlcHRoID0gMSwKc3RhdHVzID0gImFjdGl2ZSIsCnVzZXJOYW1lID0gImNyYXktbGlic2NpIiwKd1YgPSAiXjAwMDAwMDIyLjAwMDAwMDAxMi4wMDAwMDAwMDEuMDAwMDAwMDAxLip6ZmluYWwiLAp9LApbImNyYXktbXBpY2giXSA9IHsKZm4gPSAiL29wdC9jcmF5L3BlL2xtb2QvbW9kdWxlZmlsZXMv +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696887502 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier08191 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier08191 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2504.frontier08191 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r1 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465528 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465528 +_ModuleTable014_=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 +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r10.1465527.txt b/results/env.r10.1465527.txt new file mode 100644 index 0000000..abf83ee --- /dev/null +++ b/results/env.r10.1465527.txt @@ -0,0 +1,353 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=18914 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=cmRlciA9IDE2LApwcm9wVCA9IHt9LApzdGFja0RlcHRoID0gMSwKc3RhdHVzID0gImFjdGl2ZSIsCnVzZXJOYW1lID0gIkRlZkFwcHMiLAp3ViA9ICIqZGVmYXVsdC4qemZpbmFsIiwKfSwKWyJQcmdFbnYtYW1kIl0gPSB7CmZuID0gIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NvcmUvUHJnRW52LWFtZC84LjMuMy5sdWEiLApmdWxsTmFtZSA9ICJQcmdFbnYtYW1kLzguMy4zIiwKbG9hZE9yZGVyID0gMTMsCnByb3BUID0ge30sCnN0YWNrRGVwdGggPSAxLApzdGF0dXMgPSAiYWN0aXZlIiwKdXNlck5hbWUgPSAiUHJnRW52LWFtZC84LjMuMyIsCndWID0gIl4wMDAwMDAwOC4wMDAwMDAwMDMuMDAwMDAwMDAzLip6ZmluYWwiLAp9LAphbWQgPSB7CmZuID0gIi9vcHQv +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier08319 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier08319 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696894702 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=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 +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696887502 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier08319 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier08319 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2505.frontier08319 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r10 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465527 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465527 +_ModuleTable014_=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 +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r11.1465787.txt b/results/env.r11.1465787.txt new file mode 100644 index 0000000..12468d7 --- /dev/null +++ b/results/env.r11.1465787.txt @@ -0,0 +1,354 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=130390 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier07630 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +OLDPWD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier07630 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696904467 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=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 +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696897267 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier07630 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier07630 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2500.frontier07630 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r11 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=dGggPSAyLApzdGF0dXMgPSAiYWN0aXZlIiwKdXNlck5hbWUgPSAiaHNpL2RlZmF1bHQiLAp3ViA9ICIqZGVmYXVsdC4qemZpbmFsIiwKfSwKbGliZmFicmljID0gewpmbiA9ICIvb3B0L2NyYXkvbW9kdWxlZmlsZXMvbGliZmFicmljLzEuMTUuMi4wIiwKZnVsbE5hbWUgPSAibGliZmFicmljLzEuMTUuMi4wIiwKbG9hZE9yZGVyID0gMiwKcHJvcFQgPSB7fSwKc3RhY2tEZXB0aCA9IDEsCnN0YXR1cyA9ICJhY3RpdmUiLAp1c2VyTmFtZSA9ICJsaWJmYWJyaWMiLAp3ViA9ICJeMDAwMDAwMDEuMDAwMDAwMDE1LjAwMDAwMDAwMi4qemZpbmFsIiwKfSwKWyJwZXJmdG9vbHMtYmFzZSJdID0gewpmbiA9ICIvb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlL3BlcmZ0b29s +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465787 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=Y3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NvcmUvYW1kLzUuMy4wLmx1YSIsCmZ1bGxOYW1lID0gImFtZC81LjMuMCIsCmxvYWRPcmRlciA9IDgsCnByb3BUID0ge30sCnN0YWNrRGVwdGggPSAxLApzdGF0dXMgPSAiYWN0aXZlIiwKdXNlck5hbWUgPSAiYW1kIiwKd1YgPSAiXjAwMDAwMDA1LjAwMDAwMDAwMy4qemZpbmFsIiwKfSwKWyJjcmF5LWRzbW1sIl0gPSB7CmZuID0gIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NvcmUvY3JheS1kc21tbC8wLjIuMi5sdWEiLApmdWxsTmFtZSA9ICJjcmF5LWRzbW1sLzAuMi4yIiwKbG9hZE9yZGVyID0gMTAsCnByb3BUID0ge30sCnN0YWNrRGVwdGggPSAxLApzdGF0dXMgPSAiYWN0aXZlIiwKdXNlck5hbWUgPSAiY3JheS1k +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465787 +_ModuleTable014_=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 +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r12.1467862.txt b/results/env.r12.1467862.txt new file mode 100644 index 0000000..8e02512 --- /dev/null +++ b/results/env.r12.1467862.txt @@ -0,0 +1,354 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=103126 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/ccs/home/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=198.206.223.251 27061 128.219.135.137 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier00741 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +OLDPWD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/19 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier00741 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696965183 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=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 +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1643 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=ICJeMDAwMDAwMDIuMDAwMDAwMDA2LjAwMDAwMDAwMi4qemZpbmFsLS4wMDAwMDAwMDIuMDAwMDAwMDA1LipfLjAwMDAwMDAwMi4wMDAwMDAwMjIuKl8uKl8uKmdkLjAwMDAwMDA2Ny4qYy4wMDAwMDAwMDMuKmYuKnNoYXN0YS4qemZpbmFsIiwKfSwKfSwKbXBhdGhBID0gewoKIi9zdy9mcm9udGllci9zcGFjay1lbnZzL2Jhc2UvbW9kdWxlcy9zcGFjay9jcmF5LXNsZXMxNS14ODZfNjQvY3JheS1tcGljaC84LjEuMjMtenV6dHlmcS9jbGFuZy8xNS4wLjAtcm9jbTUuMy4wIgosICIvc3cvZnJvbnRpZXIvc3BhY2stZW52cy9iYXNlL21vZHVsZXMvc3BhY2svY3JheS1zbGVzMTUteDg2XzY0L2NsYW5nLzE1LjAuMC1yb2NtNS4zLjAiCiwgIi9zdy9mcm9udGllci9zcGFjay1lbnZz +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696957983 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier00741 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier00741 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login07 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=198.206.223.251 27061 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2005.frontier00741 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-9d655ef8-fa71-45eb-ab06-7f0f30b9ea29.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r12 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1467862 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=L2Jhc2UvbW9kdWxlcy9zcGFjay9saW51eC1zbGVzMTUteDg2XzY0L0NvcmUiCiwgIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL21waS9hbWQvNC4wL29maS8xLjAvY3JheS1tcGljaC84LjAiCiwgIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NvbW5ldC9hbWQvNC4wL29maS8xLjAiCiwgIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NvbXBpbGVyL2FtZC80LjAiCiwgIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL21peF9jb21waWxlcnMiCiwgIi9zdy9mcm9udGllci9zcGFjay1lbnZzL2Jhc2UvbW9kdWxlcy9zcGFjay9jcmF5LXNsZXMxNS14ODZfNjQvY3JheS1tcGljaC84LjEuMjMtYTcybXg0Mi9jY2UvMTUuMC4wIgosICIvc3cvZnJv +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=ZHVsZWZpbGVzL0NvcmU6L29wdC9jcmF5L3BlL2xtb2QvbW9kdWxlZmlsZXMvY29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jcmF5cGUtdGFyZ2V0cy9kZWZhdWx0Oi9zdy9mcm9udGllci9tb2R1bGVmaWxlczovb3B0L2NyYXkvbW9kdWxlZmlsZXMiLAp9Cg== +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=ZXMxNS14ODZfNjQvY3JheS1tcGljaC84LjEuMjMtajU2YXp3NS9jY2UvMTUuMC4wL2RhcnNoYW4tcnVudGltZS8zLjQuMC5sdWEiLApmdWxsTmFtZSA9ICJkYXJzaGFuLXJ1bnRpbWUvMy40LjAiLApsb2FkT3JkZXIgPSAxNCwKcHJvcFQgPSB7fSwKc3RhY2tEZXB0aCA9IDIsCnN0YXR1cyA9ICJhY3RpdmUiLAp1c2VyTmFtZSA9ICJkYXJzaGFuLXJ1bnRpbWUiLAp3ViA9ICIwMDAwMDAwMDMuMDAwMDAwMDA0Lip6ZmluYWwiLAp9LApoc2kgPSB7CmZuID0gIi9zdy9mcm9udGllci9tb2R1bGVmaWxlcy9oc2kvZGVmYXVsdC5sdWEiLApmdWxsTmFtZSA9ICJoc2kvZGVmYXVsdCIsCmxvYWRPcmRlciA9IDE1LApwcm9wVCA9IHt9LApzdGFja0RlcHRoID0gMiwKc3RhdHVzID0gImFj +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login07 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/ccs/home/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1467862 +_ModuleTable014_=bnRpZXIvc3BhY2stZW52cy9iYXNlL21vZHVsZXMvc3BhY2svY3JheS1zbGVzMTUteDg2XzY0L2NyYXktbXBpY2gvOC4xLjIzLWo1NmF6dzUvY2NlLzE1LjAuMCIKLCAiL3N3L2Zyb250aWVyL3NwYWNrLWVudnMvYmFzZS9tb2R1bGVzL3NwYWNrL2NyYXktc2xlczE1LXg4Nl82NC9jY2UvMTUuMC4wIgosICIvb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9tcGkvY3JheWNsYW5nLzE0LjAvb2ZpLzEuMC9jcmF5LW1waWNoLzguMCIKLCAiL29wdC9jcmF5L3BlL2xtb2QvbW9kdWxlZmlsZXMvcGVyZnRvb2xzLzIyLjEyLjAiCiwgIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL25ldC9vZmkvMS4wIgosICIvb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jcHUveDg2 +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r2.1465529.txt b/results/env.r2.1465529.txt new file mode 100644 index 0000000..e3b13ca --- /dev/null +++ b/results/env.r2.1465529.txt @@ -0,0 +1,353 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=66969 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier08320 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier08320 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696894702 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=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 +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696887502 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier08320 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier08320 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2505.frontier08320 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r2 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465529 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465529 +_ModuleTable014_=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 +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r3.1465530.txt b/results/env.r3.1465530.txt new file mode 100644 index 0000000..5b045e4 --- /dev/null +++ b/results/env.r3.1465530.txt @@ -0,0 +1,353 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=60998 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier08444 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier08444 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696894702 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=c21tbCIsCndWID0gIl4wMDAwMDAwMC4wMDAwMDAwMDIuMDAwMDAwMDAyLip6ZmluYWwiLAp9LApbImNyYXktbGlic2NpIl0gPSB7CmZuID0gIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NvcmUvY3JheS1saWJzY2kvMjIuMTIuMS4xLmx1YSIsCmZ1bGxOYW1lID0gImNyYXktbGlic2NpLzIyLjEyLjEuMSIsCmxvYWRPcmRlciA9IDEyLApwcm9wVCA9IHt9LApzdGFja0RlcHRoID0gMSwKc3RhdHVzID0gImFjdGl2ZSIsCnVzZXJOYW1lID0gImNyYXktbGlic2NpIiwKd1YgPSAiXjAwMDAwMDIyLjAwMDAwMDAxMi4wMDAwMDAwMDEuMDAwMDAwMDAxLip6ZmluYWwiLAp9LApbImNyYXktbXBpY2giXSA9IHsKZm4gPSAiL29wdC9jcmF5L3BlL2xtb2QvbW9kdWxlZmlsZXMv +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696887502 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier08444 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier08444 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2506.frontier08444 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r3 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465530 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465530 +_ModuleTable014_=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 +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r4.1465531.txt b/results/env.r4.1465531.txt new file mode 100644 index 0000000..14317dc --- /dev/null +++ b/results/env.r4.1465531.txt @@ -0,0 +1,353 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=44175 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=cyA9ICJhY3RpdmUiLAp1c2VyTmFtZSA9ICJjcmF5cGUtbmV0d29yay1vZmkiLAp3ViA9ICJNLip6ZmluYWwiLAp9LApbImNyYXlwZS14ODYtdHJlbnRvIl0gPSB7CmZuID0gIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQvY3JheXBlLXg4Ni10cmVudG8ubHVhIiwKZnVsbE5hbWUgPSAiY3JheXBlLXg4Ni10cmVudG8iLApsb2FkT3JkZXIgPSAxLApwcm9wVCA9IHt9LApzdGFja0RlcHRoID0gMCwKc3RhdHVzID0gImFjdGl2ZSIsCnVzZXJOYW1lID0gImNyYXlwZS14ODYtdHJlbnRvIiwKd1YgPSAiTS4qemZpbmFsIiwKfSwKWyJkYXJzaGFuLXJ1bnRpbWUiXSA9IHsKZm4gPSAiL3N3L2Zyb250aWVyL3NwYWNrLWVudnMvYmFzZS9t +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier07674 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier07674 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696895063 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=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 +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696887863 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier07674 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier07674 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2500.frontier07674 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r4 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465531 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=LAp1c2VyTmFtZSA9ICJjcmF5LXBtaSIsCndWID0gIl4wMDAwMDAwNi4wMDAwMDAwMDEuMDAwMDAwMDA4Lip6ZmluYWwiLAp9LApbImNyYXktcHl0aG9uIl0gPSB7CmZuID0gIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NvcmUvY3JheS1weXRob24vMy45LjEzLjEubHVhIiwKZnVsbE5hbWUgPSAiY3JheS1weXRob24vMy45LjEzLjEiLApsb2FkT3JkZXIgPSA3LApwcm9wVCA9IHt9LApzdGFja0RlcHRoID0gMCwKc3RhdHVzID0gImFjdGl2ZSIsCnVzZXJOYW1lID0gImNyYXktcHl0aG9uIiwKd1YgPSAiXjAwMDAwMDAzLjAwMDAwMDAwOS4wMDAwMDAwMTMuMDAwMDAwMDAxLip6ZmluYWwiLAp9LApjcmF5cGUgPSB7CmZuID0gIi9vcHQvY3JheS9wZS9sbW9kL21vZHVs +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465531 +_ModuleTable014_=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 +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r5.1465786.txt b/results/env.r5.1465786.txt new file mode 100644 index 0000000..9e3bee8 --- /dev/null +++ b/results/env.r5.1465786.txt @@ -0,0 +1,354 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=20456 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier03063 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +OLDPWD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier03063 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696903920 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=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 +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696896720 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier03063 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier03063 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2111.frontier03063 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r5 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465786 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465786 +_ModuleTable014_=ZS8xNS4wLjAiCiwgIi9zdy9mcm9udGllci9zcGFjay1lbnZzL2Jhc2UvbW9kdWxlcy9zcGFjay9jcmF5LXNsZXMxNS14ODZfNjQvY3JheS1tcGljaC84LjEuMjMtajU2YXp3NS9jY2UvMTUuMC4wIgosICIvc3cvZnJvbnRpZXIvc3BhY2stZW52cy9iYXNlL21vZHVsZXMvc3BhY2svY3JheS1zbGVzMTUteDg2XzY0L2NjZS8xNS4wLjAiCiwgIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL21waS9jcmF5Y2xhbmcvMTQuMC9vZmkvMS4wL2NyYXktbXBpY2gvOC4wIgosICIvb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9wZXJmdG9vbHMvMjIuMTIuMCIKLCAiL29wdC9jcmF5L3BlL2xtb2QvbW9kdWxlZmlsZXMvbmV0L29maS8xLjAiCiwgIi9vcHQvY3JheS9wZS9sbW9k +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r6.1465533.txt b/results/env.r6.1465533.txt new file mode 100644 index 0000000..b66efdf --- /dev/null +++ b/results/env.r6.1465533.txt @@ -0,0 +1,353 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=1526 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier07936 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier07936 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696895063 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=c21tbCIsCndWID0gIl4wMDAwMDAwMC4wMDAwMDAwMDIuMDAwMDAwMDAyLip6ZmluYWwiLAp9LApbImNyYXktbGlic2NpIl0gPSB7CmZuID0gIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NvcmUvY3JheS1saWJzY2kvMjIuMTIuMS4xLmx1YSIsCmZ1bGxOYW1lID0gImNyYXktbGlic2NpLzIyLjEyLjEuMSIsCmxvYWRPcmRlciA9IDEyLApwcm9wVCA9IHt9LApzdGFja0RlcHRoID0gMSwKc3RhdHVzID0gImFjdGl2ZSIsCnVzZXJOYW1lID0gImNyYXktbGlic2NpIiwKd1YgPSAiXjAwMDAwMDIyLjAwMDAwMDAxMi4wMDAwMDAwMDEuMDAwMDAwMDAxLip6ZmluYWwiLAp9LApbImNyYXktbXBpY2giXSA9IHsKZm4gPSAiL29wdC9jcmF5L3BlL2xtb2QvbW9kdWxlZmlsZXMv +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696887863 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier07936 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier07936 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2502.frontier07936 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r6 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465533 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465533 +_ModuleTable014_=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 +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r7.1465534.txt b/results/env.r7.1465534.txt new file mode 100644 index 0000000..77fdc8f --- /dev/null +++ b/results/env.r7.1465534.txt @@ -0,0 +1,353 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=14667 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier08051 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier08051 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696895063 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=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 +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696887863 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier08051 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier08051 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2503.frontier08051 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r7 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465534 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465534 +_ModuleTable014_=ZS8xNS4wLjAiCiwgIi9zdy9mcm9udGllci9zcGFjay1lbnZzL2Jhc2UvbW9kdWxlcy9zcGFjay9jcmF5LXNsZXMxNS14ODZfNjQvY3JheS1tcGljaC84LjEuMjMtajU2YXp3NS9jY2UvMTUuMC4wIgosICIvc3cvZnJvbnRpZXIvc3BhY2stZW52cy9iYXNlL21vZHVsZXMvc3BhY2svY3JheS1zbGVzMTUteDg2XzY0L2NjZS8xNS4wLjAiCiwgIi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL21waS9jcmF5Y2xhbmcvMTQuMC9vZmkvMS4wL2NyYXktbXBpY2gvOC4wIgosICIvb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9wZXJmdG9vbHMvMjIuMTIuMCIKLCAiL29wdC9jcmF5L3BlL2xtb2QvbW9kdWxlZmlsZXMvbmV0L29maS8xLjAiCiwgIi9vcHQvY3JheS9wZS9sbW9k +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r8.1465535.txt b/results/env.r8.1465535.txt new file mode 100644 index 0000000..54d17d1 --- /dev/null +++ b/results/env.r8.1465535.txt @@ -0,0 +1,353 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=10201 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier10439 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier10439 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696895065 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=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 +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696887865 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier10439 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier10439 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2611.frontier10439 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r8 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465535 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465535 +_ModuleTable014_=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 +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/env.r9.1465536.txt b/results/env.r9.1465536.txt new file mode 100644 index 0000000..6d6bda7 --- /dev/null +++ b/results/env.r9.1465536.txt @@ -0,0 +1,353 @@ +__LMOD_REF_COUNT_PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH:1 +CRAY_LMOD_MPI=cray-mpich/8.0 +SLURM_MPI_TYPE=cray_shasta +LMOD_FAMILY_CRAYPE_NETWORK_VERSION=false +LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:/opt/rocm-5.3.0/hsa/lib:/opt/rocm-5.3.0/lib64:/opt/rocm-5.3.0/lib:/opt/rocm-5.3.0/llvm/lib:/opt/cray/pe/python/3.9.13.1/lib:/opt/cray/pe/gcc-libs:/opt/cray/pe/papi/6.0.0.17/lib64:/opt/cray/libfabric/1.15.2.0/lib64 +LS_COLORS=no=00:fi=00:di=01;34:ln=00;36:pi=40;33:so=01;35:do=01;35:bd=40;33;01:cd=40;33;01:or=41;33;01:ex=00;32:*.cmd=00;32:*.exe=01;32:*.com=01;32:*.bat=01;32:*.btm=01;32:*.dll=01;32:*.tar=00;31:*.tbz=00;31:*.tgz=00;31:*.rpm=00;31:*.deb=00;31:*.arj=00;31:*.taz=00;31:*.lzh=00;31:*.lzma=00;31:*.zip=00;31:*.zoo=00;31:*.z=00;31:*.Z=00;31:*.gz=00;31:*.bz2=00;31:*.tb2=00;31:*.tz2=00;31:*.tbz2=00;31:*.xz=00;31:*.avi=01;35:*.bmp=01;35:*.dl=01;35:*.fli=01;35:*.gif=01;35:*.gl=01;35:*.jpg=01;35:*.jpeg=01;35:*.mkv=01;35:*.mng=01;35:*.mov=01;35:*.mp4=01;35:*.mpg=01;35:*.pcx=01;35:*.pbm=01;35:*.pgm=01;35:*.png=01;35:*.ppm=01;35:*.svg=01;35:*.tga=01;35:*.tif=01;35:*.webm=01;35:*.webp=01;35:*.wmv=01;35:*.xbm=01;35:*.xcf=01;35:*.xpm=01;35:*.aiff=00;32:*.ape=00;32:*.au=00;32:*.flac=00;32:*.m4a=00;32:*.mid=00;32:*.mp3=00;32:*.mpc=00;32:*.ogg=00;32:*.voc=00;32:*.wav=00;32:*.wma=00;32:*.wv=00;32: +__LMOD_REF_COUNT_PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig:1 +HOSTTYPE=x86_64 +CRAY_DSMML_ROOTDIR=/opt/cray/pe/dsmml/0.2.2 +SLURM_NODEID=0 +CRAY_DSMML_DIR=/opt/cray/pe/dsmml/0.2.2/dsmml +SLURM_TASK_PID=53951 +__LMOD_REF_COUNT_PATH=/sw/sources/hpss/bin:1;/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:2;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:1;/opt/cray/pe/mpich/8.1.23/bin:1;/opt/cray/pe/craype/2.7.19/bin:1;/opt/rocm-5.3.0/bin:1;/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:1;/opt/cray/pe/python/3.9.13.1/bin:1;/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:1;/opt/cray/pe/perftools/22.12.0/bin:1;/opt/cray/pe/papi/6.0.0.17/bin:1;/opt/cray/libfabric/1.15.2.0/bin:1;/opt/clmgr/sbin:1;/opt/clmgr/bin:1;/opt/sgi/sbin:1;/opt/sgi/bin:1;/sw/frontier/bin:1;/usr/local/bin:1;/usr/bin:1;/bin:2;/opt/bin:1;/opt/c3/bin:1;/usr/lib/mit/bin:1;/opt/puppetlabs/bin:1;/sbin:1;/opt/cray/pe/bin:1 +_ModuleTable002_=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 +PE_MPICH_GTL_DIR_amd_gfx90a=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_OMP_REQUIRES= +SSH_CONNECTION=134.253.242.223 59145 128.219.135.142 22 +LESSCLOSE=lessclose.sh %s %s +CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:/opt/cray/pe/mpich/8.1.23/gtl/lib:/opt/cray/pe/dsmml/0.2.2/dsmml/lib:/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:/opt/cray/pe/pmi/6.1.8/lib:/opt/cray/pe/perftools/22.12.0/lib64 +SLURM_PRIO_PROCESS=0 +XKEYSYMDB=/usr/X11R6/lib/X11/XKeysymDB +PE_ENV=AMD +__LMOD_REF_COUNT_PE_DSMML_PKGCONFIG_LIBS=dsmml:1 +__LMOD_REF_COUNT_C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +__LMOD_REF_COUNT_PE_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +LMOD_FAMILY_CRAYPE_CPU=craype-x86-trento +CRAYPAT_LD_LIBRARY_PATH=/opt/cray/pe/gcc-libs:/opt/cray/gcc-libs:/opt/cray/pe/perftools/22.12.0/lib64 +CRAY_LMOD_COMPILER=amd/4.0 +LMOD_MODULERCFILE=/sw/frontier/lmod/etc/rc.lua +CRAY_DSMML_PREFIX=/opt/cray/pe/dsmml/0.2.2/dsmml +PE_PRODUCT_LIST=CRAY_LIBSCI:CRAYPE:CRAY_PMI:CRAYPE_X86_TRENTO:PERFTOOLS:CRAYPAT +LMOD_FAMILY_MPI_VERSION=8.1.23 +CRAYPAT_ROOT=/opt/cray/pe/perftools/22.12.0 +PE_PERFTOOLS_MPICH_LIBDIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib +PE_MPICH_GTL_DIR_amd_gfx906=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +PE_LIBSCI_MODULE_NAME=cray-libsci/22.12.1.1 +LANG=en_US.UTF-8 +PE_LIBSCI_GENCOMPILERS_CRAY_x86_64=9.0 +LMOD_SYSTEM_NAME=frontier +SLURM_SUBMIT_DIR=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth +PE_MPICH_GTL_DIR_amd_gfx908=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +MEMBERWORK=/lustre/orion/scratch/cpearson +PE_DSMML_MODULE_NAME=cray-dsmml +WINDOWMANAGER=xterm +_ModuleTable008_=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 +HIP_LIB_PATH=/opt/rocm-5.3.0/lib +LESS=-M -I -R +PE_MPICH_GTL_LIBS_nvidia70=-lmpi_gtl_cuda +JAVA_ROOT=/usr/lib64/jvm/java +LMOD_FAMILY_COMPILER_VERSION=5.3.0 +HOSTNAME=frontier05673 +LMOD_SYSTEM_DEFAULT_MODULES=craype-x86-trento:craype-network-ofi:perftools-base:xpmem:cray-pmi:PrgEnv-cray:DefApps +CSHEDIT=emacs +C_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +ENVIRONMENT=BATCH +GPG_TTY=/dev/pts/21 +MPICH_OFI_NIC_POLICY=NUMA +LESS_ADVANCED_PREPROCESSOR=no +COLORTERM=truecolor +OLCF_DARSHAN_RUNTIME_ROOT=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp +_ModuleTable007_=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 +CRAY_PERFTOOLS_VERSION=22.12.0 +ROCR_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 +ROCM_PATH=/opt/rocm-5.3.0 +__LMOD_REF_COUNT_CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include:1 +LMOD_FAMILY_CRAYPE_NETWORK=craype-network-ofi +JAVA_HOME=/usr/lib64/jvm/java +LMOD_FAMILY_CRAYPE_VERSION=2.7.19 +VSCODE_GIT_ASKPASS_EXTRA_ARGS= +PE_XPMEM_PKGCONFIG_LIBS=cray-xpmem +SLURM_PROCID=0 +APP2_STATE=22.12.0 +SLURM_JOB_GID=27732 +MACHTYPE=x86_64-suse-linux +PE_LIBSCI_FIXED_PRGENV=CRAYCLANG +__LMOD_REF_COUNT_LD_LIBRARY_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib:1;/opt/rocm-5.3.0/hsa/lib:1;/opt/rocm-5.3.0/lib64:1;/opt/rocm-5.3.0/lib:1;/opt/rocm-5.3.0/llvm/lib:1;/opt/cray/pe/python/3.9.13.1/lib:1;/opt/cray/pe/gcc-libs:1;/opt/cray/pe/papi/6.0.0.17/lib64:1;/opt/cray/libfabric/1.15.2.0/lib64:1 +__LMOD_REF_COUNT_PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:2;/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:1;/opt/cray/pe/craype/2.7.19/pkg-config:1;/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:1;/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig:1 +SLURMD_NODENAME=frontier05673 +PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:PE_MPICH:PE_DSMML:PE_PMI:PE_XPMEM +_ModuleTable015_=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 +SLURM_JOB_END_TIME=1696895123 +CRAY_MPICH_BASEDIR=/opt/cray/pe/mpich/8.1.23/ofi +PE_MPICH_GTL_DIR_nvidia80=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +_ModuleTable004_=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 +VIRTUAL_ENV=/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv +PE_LIBSCI_GENCOMPILERS_GNU_x86_64=9.1 +MINICOM=-c on +SLURM_TASKS_PER_NODE=1 +PAT_BUILD_PAPI_LIBDIR=/opt/cray/pe/papi/6.0.0.17/lib64 +__LMOD_REF_COUNT_PYTHONPATH=/opt/cray/pe/python/3.9.13.1:1 +SCOPE_BUILD=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build +C3_RSH=ssh -oConnectTimeout=10 -oForwardX11=no +PE_MPICH_PKGCONFIG_VARIABLES=PE_MPICH_GTL_DIR_@accelerator@:PE_MPICH_GTL_LIBS_@accelerator@ +PE_MPICH_PKGCONFIG_LIBS=mpich +QT_SYSTEM_DIR=/usr/share/desktop-data +OSTYPE=linux +OLCF_FAMILY_CRAYPE=craype +XDG_SESSION_ID=1115 +CRAY_DSMML_VER=0.2.2 +__LMOD_REF_COUNT_PYTHON_PATH=/opt/cray/pe/python/3.9.13.1:1 +SLURM_NNODES=1 +USER=cpearson +PAGER=less +RFE_811452_DISABLE=1 +CRAY_PYTHON_LEVEL=3.9.13.1 +LMOD_FAMILY_PRGENV_VERSION=8.3.3 +ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +OLCF_FAMILY_CRAYPE_NETWORK_VERSION=false +_ModuleTable012_=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 +CRAY_CPU_TARGET=x86-trento +SLURM_JOB_START_TIME=1696887923 +PE_LIBSCI_VOLATILE_PRGENV=AMD CRAY GNU +PKGCONFIG_ENABLED=1 +__LMOD_REF_COUNT_MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:1;/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:1;/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:1;/opt/cray/pe/lmod/modulefiles/mix_compilers:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:1;/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:1;/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:1;/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:1;/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:1;/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:1;/opt/cray/pe/modulefiles/Linux:1;/opt/cray/pe/modulefiles/Core:1;/opt/cray/pe/lmod/lmod/modulefiles/Core:1;/opt/cray/pe/lmod/modulefiles/core:1;/opt/cray/pe/lmod/modulefiles/craype-targets/default:1;/sw/frontier/modulefiles:1;/opt/cray/modulefiles:1 +MORE=-sl +CRAY_PERFTOOLS_PREFIX=/opt/cray/pe/perftools/22.12.0 +PE_FORTRAN_PKGCONFIG_LIBS=mpichf90 +__LMOD_REF_COUNT_CRAY_LD_LIBRARY_PATH=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64/lib:1;/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib:1;/opt/cray/pe/mpich/8.1.23/gtl/lib:1;/opt/cray/pe/dsmml/0.2.2/dsmml/lib:1;/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib:1;/opt/cray/pe/pmi/6.1.8/lib:1;/opt/cray/pe/perftools/22.12.0/lib64:1 +OLCF_FAMILY_MPI_VERSION=8.1.23 +CRAY_LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +PWD=/ccs/home/cpearson/proj/frontier-gpu-bandwidth +OLCF_FAMILY_COMPILER_VERSION=5.3.0 +PE_MPICH_GTL_DIR_nvidia70=-L/opt/cray/pe/mpich/8.1.23/gtl/lib +SLURM_JOB_NODELIST=frontier05673 +HOME=/ccs/home/cpearson +OLCF_FAMILY_CRAYPE_CPU=craype-x86-trento +SLURM_CLUSTER_NAME=frontier +CRAY_PMI_INCLUDE_OPTS=-I/opt/cray/pe/pmi/6.1.8/include +PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig +PELOCAL_PRGENV=true +BROWSER=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/helpers/browser.sh +VSCODE_GIT_ASKPASS_NODE=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/node +CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:/opt/rocm-5.3.0/hip:/opt/rocm-5.3.0 +PE_LIBSCI_REQUIRED_PRODUCTS=PE_MPICH +SLURM_NODELIST=frontier05673 +PE_LIBSCI_GENCOMPS_AMD_x86_64=40 +TERM_PROGRAM=vscode +SLURM_GPUS_ON_NODE=8 +HOST=login12 +__LMOD_REF_COUNT_PE_PMI_PKGCONFIG_LIBS=cray-pmi:1 +SSH_CLIENT=134.253.242.223 59145 22 +TERM_PROGRAM_VERSION=1.82.3 +LMOD_VERSION=8.7.15 +XNLSPATH=/usr/share/X11/nls +PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90 +SLURM_NTASKS=1 +PE_LIBSCI_PKGCONFIG_LIBS=libsci_mpi:libsci +XDG_SESSION_TYPE=tty +SCOPE_RESULTS=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/run +SDK_HOME=/usr/lib64/jvm/java +CRAY_LMOD_CPU=x86-trento/1.0 +PE_MPICH_GTL_LIBS_nvidia80=-lmpi_gtl_cuda +PE_LIBSCI_OMP_REQUIRES_openmp=_mp +LMOD_FAMILY_CRAYPE_CPU_VERSION=false +SLURM_JOB_CPUS_PER_NODE=56 +BASH_ENV=/opt/cray/pe/lmod/lmod/init/bash +XDG_DATA_DIRS=/usr/share +__LMOD_REF_COUNT_PE_MPICH_FORTRAN_PKGCONFIG_LIBS=mpichf90:1 +SLURM_TOPOLOGY_ADDR=root.s2309.frontier05673 +SLURM_THREADS_PER_CORE=1 +PYTHON_PATH=/opt/cray/pe/python/3.9.13.1 +CRAY_LMOD_NET=ofi/1.0 +PE_PMI_PKGCONFIG_LIBS=cray-pmi +__LMOD_REF_COUNT_PE_PKGCONFIG_PRODUCTS=PE_LIBSCI:1;PE_MPICH:1;PE_DSMML:1;PE_PMI:1;PE_XPMEM:1 +PE_MPICH_MODULE_NAME=cray-mpich +OLCF_FAMILY_MPI=cray-mpich +CRAY_LIBSCI_DIR=/opt/cray/pe/libsci/22.12.1.1 +LIBGL_DEBUG=quiet +SLURM_WORKING_CLUSTER=frontier:slurm1.frontier.olcf.ornl.gov:6817:9984:109 +PE_LIBSCI_VOLATILE_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/@PRGENV@/@PE_LIBSCI_GENCOMPS@/@PE_LIBSCI_TARGET@/lib/pkgconfig +JDK_HOME=/usr/lib64/jvm/java +__LMOD_REF_COUNT_PE_MPICH_FIXED_PRGENV=AMD:1 +OLCF_FAMILY_CRAYPE_VERSION=2.7.19 +LMOD_SHELL_PRGM=bash +__LMOD_REF_COUNT_LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib:1 +VSCODE_IPC_HOOK_CLI=/run/user/15925/vscode-ipc-256d0a63-d03f-4e14-9f9b-a9fe229fd587.sock +LIBSCI_VERSION=22.12.1.1 +SLURM_JOB_NAME=r9 +PROFILEREAD=true +TMPDIR=/tmp +LIBRARY_PATH=/opt/rocm-5.3.0/llvm/lib +HSA_PATH=/opt/rocm-5.3.0/hsa +PERFTOOLS_VERSION=22.12.0 +SLURM_JOB_GPUS=0,1,2,3,4,5,6,7 +__LMOD_REF_COUNT_CMAKE_PREFIX_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp:2;/opt/rocm-5.3.0/hip:2;/opt/rocm-5.3.0:2 +LMOD_sys=Linux +_ModuleTable010_=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 +PE_PKGCONFIG_LIBS=darshan-runtime:libsci_mpi:libsci:mpich:dsmml +CRAY_PMI_PREFIX=/opt/cray/pe/pmi/6.1.8 +SLURM_JOBID=1465536 +_ModuleTable001_=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 +SLURM_CONF=/var/spool/slurmd/conf-cache/slurm.conf +CRAY_AMD_COMPILER_PREFIX=/opt/rocm-5.3.0 +LOADEDMODULES=craype-x86-trento:libfabric/1.15.2.0:craype-network-ofi:perftools-base/22.12.0:xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:cray-pmi/6.1.8:cray-python/3.9.13.1:amd/5.3.0:craype/2.7.19:cray-dsmml/0.2.2:cray-mpich/8.1.23:cray-libsci/22.12.1.1:PrgEnv-amd/8.3.3:darshan-runtime/3.4.0:hsi/default:DefApps/default +CRAY_AMD_COMPILER_VERSION=5.3.0 +OLCF_FAMILY_CRAYPE_CPU_VERSION=false +CRAYPE_NETWORK_TARGET=ofi +_ModuleTable006_=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 +__LMOD_REF_COUNT_MANPATH=/sw/sources/hpss/man:1;/opt/cray/pe/libsci/22.12.1.1/man:1;/opt/cray/pe/man/csmlversion:1;/opt/cray/pe/mpich/8.1.23/ofi/man:1;/opt/cray/pe/mpich/8.1.23/man/mpich:1;/opt/cray/pe/dsmml/0.2.2/dsmml/man:1;/opt/cray/pe/craype/2.7.19/man:1;/opt/cray/pe/python/3.9.13.1/share/man:1;/opt/cray/pe/pmi/6.1.8/man:1;/opt/cray/pe/perftools/22.12.0/man:1;/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:1;/opt/cray/libfabric/1.15.2.0/share/man:1;/opt/cray/pe/lmod/lmod/share/man:1;/usr/local/man:1;/usr/share/man:1;/usr/man:1;/opt/c3/man:1;/opt/puppetlabs/puppet/share/man:1;/opt/clmgr/man:1;/opt/sgi/share/man:1;/opt/clmgr/share/man:1;/opt/clmgr/lib/cm-cli/man:1 +_ModuleTable003_=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 +SLURM_NODE_ALIASES=(null) +SLURM_JOB_QOS=normal +LMOD_ROOT=/opt/cray/pe/lmod +SLURM_TOPOLOGY_ADDR_PATTERN=switch.switch.node +PAT_RT_PERFCTR_DISABLE_COMPONENTS=nvml,cuda +PE_CRAYCLANG_FIXED_PKGCONFIG_PATH=/opt/cray/pe/libsci/22.12.1.1/CRAY/9.0/x86_64/lib/pkgconfig +FROM_HEADER= +CRAY_MPICH_ROOTDIR=/opt/cray/pe/mpich/8.1.23 +_ModuleTable013_=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 +MAIL=/var/spool/mail/cpearson +SLURM_CPUS_ON_NODE=56 +SLURM_SPANK_GPU_POWER_CAP_SRC=default +VSCODE_GIT_ASKPASS_MAIN=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass-main.js +SLURM_JOB_NUM_NODES=1 +_ModuleTable016_=Y3JheS9wZS9sbW9kL2xtb2QvbW9kdWxlZmlsZXMvQ29yZTovb3B0L2NyYXkvcGUvbG1vZC9tb2R1bGVmaWxlcy9jb3JlOi9vcHQvY3JheS9wZS9sbW9kL21vZHVsZWZpbGVzL2NyYXlwZS10YXJnZXRzL2RlZmF1bHQ6L3N3L2Zyb250aWVyL21vZHVsZWZpbGVzOi9vcHQvY3JheS9tb2R1bGVmaWxlcyIsCn0K +LESSKEY=/etc/lesskey.bin +_ModuleTable009_=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 +SHELL=/bin/bash +TERM=xterm-256color +CRAY_MPICH_VER=8.1.23 +LMOD_ADMIN_FILE=/sw/frontier/lmod/etc/admin.list +SCOPE_SRC=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/comm_scope +XDG_SESSION_CLASS=user +PE_MPICH_FIXED_PRGENV=AMD +OLCF_FAMILY_PRGENV_VERSION=8.3.3 +LMOD_SITE_NAME=OLCF +PE_LIBSCI_GENCOMPILERS_AMD_x86_64=4.0 +LMOD_FAMILY_PRGENV=PrgEnv-amd +SLURM_JOB_UID=15925 +_ModuleTable_Sz_=16 +CRAY_LIBSCI_VERSION=22.12.1.1 +CRAY_PYTHON_PREFIX=/opt/cray/pe/python/3.9.13.1 +XCURSOR_THEME=DMZ +LS_OPTIONS=-N --color=tty -T 0 +SLURM_JOB_PARTITION=batch +ROCM_COMPILER_VERSION=5.3.0 +LMOD_FAMILY_COMPILER=amd +USE_PCM_DB=2 +CRAY_DSMML_BASEDIR=/opt/cray/pe/dsmml/0.2.2 +CRAY_XPMEM_POST_LINK_OPTS=-L/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/lib64 +__LMOD_REF_COUNT_PE_AMD_FIXED_PKGCONFIG_PATH=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/lib/pkgconfig:1 +PROJWORK=/lustre/orion/proj-shared +HWLOC_PCI_LOCALITY=/usr/share/hwloc/pci-locality-hpe-cray-ex235a +SLURM_JOB_USER=cpearson +PYTHONSTARTUP=/etc/pythonstart +SLURM_NPROCS=1 +SHLVL=6 +LANGUAGE=en_US.UTF-8 +SLURM_SUBMIT_HOST=login12 +CRAY_LIBSCI_PREFIX_DIR=/opt/cray/pe/libsci/22.12.1.1/AMD/4.0/x86_64 +G_FILENAME_ENCODING=@locale,UTF-8,ISO-8859-15,CP1252 +PYTHONPATH=/opt/cray/pe/python/3.9.13.1 +SLURM_JOB_ACCOUNT=csc465 +VSCODE_GIT_IPC_HANDLE=/run/user/15925/vscode-git-92c44bf4a7.sock +HIP_PATH=/opt/rocm-5.3.0/hip +MANPATH=/sw/sources/hpss/man:/opt/cray/pe/libsci/22.12.1.1/man:/opt/cray/pe/man/csmlversion:/opt/cray/pe/mpich/8.1.23/ofi/man:/opt/cray/pe/mpich/8.1.23/man/mpich:/opt/cray/pe/dsmml/0.2.2/dsmml/man:/opt/cray/pe/craype/2.7.19/man:/opt/cray/pe/python/3.9.13.1/share/man:/opt/cray/pe/pmi/6.1.8/man:/opt/cray/pe/perftools/22.12.0/man:/opt/cray/pe/papi/6.0.0.17/share/pdoc/man:/opt/cray/libfabric/1.15.2.0/share/man:/opt/cray/pe/lmod/lmod/share/man:/usr/local/man:/usr/share/man:/usr/man:/opt/c3/man:/opt/puppetlabs/puppet/share/man:/opt/clmgr/man:/opt/sgi/share/man:/opt/clmgr/share/man:/opt/clmgr/lib/cm-cli/man +CRAY_DSMML_VERSION=0.2.2 +SLURM_SPANK_GPU_SRANGE=500-1700 +FI_CXI_ATS=0 +MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +OSCAR_HOME=/opt/oscar +OLCF_FAMILY_PRGENV=PrgEnv-amd +MODULEPATH=/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-zuztyfq/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/clang/15.0.0-rocm5.3.0:/sw/frontier/spack-envs/base/modules/spack/linux-sles15-x86_64/Core:/opt/cray/pe/lmod/modulefiles/mpi/amd/4.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0:/opt/cray/pe/lmod/modulefiles/compiler/amd/4.0:/opt/cray/pe/lmod/modulefiles/mix_compilers:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-a72mx42/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cce/15.0.0:/opt/cray/pe/lmod/modulefiles/mpi/crayclang/14.0/ofi/1.0/cray-mpich/8.0:/opt/cray/pe/lmod/modulefiles/perftools/22.12.0:/opt/cray/pe/lmod/modulefiles/net/ofi/1.0:/opt/cray/pe/lmod/modulefiles/cpu/x86-trento/1.0:/opt/cray/pe/modulefiles/Linux:/opt/cray/pe/modulefiles/Core:/opt/cray/pe/lmod/lmod/modulefiles/Core:/opt/cray/pe/lmod/modulefiles/core:/opt/cray/pe/lmod/modulefiles/craype-targets/default:/sw/frontier/modulefiles:/opt/cray/modulefiles +CRAY_MPICH_DIR=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +SLURM_GTIDS=0 +OLCF_HSI_ROOT=/sw/sources/hpss +LMOD_FAMILY_CRAYPE=craype +LOGNAME=cpearson +DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/15925/bus +ROCM_COMPILER_PATH=/opt/rocm-5.3.0/llvm +CRAY_PMI_VERSION=6.1.8 +CRAY_MPICH_VERSION=8.1.23 +GIT_ASKPASS=/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/extensions/git/dist/askpass.sh +XDG_RUNTIME_DIR=/run/user/15925 +CPLUS_INCLUDE_PATH=/opt/rocm-5.3.0/llvm/include +MODULEPATH_ROOT=/opt/cray/pe/modulefiles +LMOD_PACKAGE_PATH=/sw/frontier/lmod/libexec +__LMOD_REF_COUNT_PE_MPICH_GENCOMPILERS_AMD=5.0:1 +PE_MPICH_GTL_LIBS_amd_gfx906=-lmpi_gtl_hsa +PE_DSMML_PKGCONFIG_LIBS=dsmml +JRE_HOME=/usr/lib64/jvm/java +PE_LIBSCI_PKGCONFIG_VARIABLES=PE_LIBSCI_OMP_REQUIRES_@openmp@:PE_SCI_EXT_LIBPATH:PE_SCI_EXT_LIBNAME +HSA_XNACK=1 +PE_LIBSCI_GENCOMPS_CRAY_x86_64=90 +XDG_CONFIG_DIRS=/etc/xdg +PATH=/sw/sources/hpss/bin:/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/bin:/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0/bin:/opt/cray/pe/mpich/8.1.23/bin:/opt/cray/pe/craype/2.7.19/bin:/opt/rocm-5.3.0/bin:/autofs/nccs-svm1_home1/cpearson/proj/frontier-gpu-bandwidth/.venv/bin:/opt/cray/pe/python/3.9.13.1/bin:/autofs/nccs-svm1_home1/cpearson/.vscode-server/bin/fdb98833154679dbaa7af67a5a29fe19e55c2b73/bin/remote-cli:/opt/cray/pe/perftools/22.12.0/bin:/opt/cray/pe/papi/6.0.0.17/bin:/opt/cray/libfabric/1.15.2.0/bin:/opt/clmgr/sbin:/opt/clmgr/bin:/opt/sgi/sbin:/opt/sgi/bin:/sw/frontier/bin:/usr/local/bin:/usr/bin:/bin:/opt/bin:/opt/c3/bin:/usr/lib/mit/bin:/opt/puppetlabs/bin:/sbin:/opt/cray/pe/bin +PE_MPICH_GTL_LIBS_amd_gfx908=-lmpi_gtl_hsa +SLURM_JOB_ID=1465536 +_ModuleTable014_=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 +JAVA_BINDIR=/usr/lib64/jvm/java/bin +__LMOD_REF_COUNT_PE_PRODUCT_LIST=CRAY_LIBSCI:1;CRAYPE:1;CRAY_PMI:1;CRAYPE_X86_TRENTO:1;PERFTOOLS:1;CRAYPAT:1 +_LMFILES_=/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-x86-trento.lua:/opt/cray/modulefiles/libfabric/1.15.2.0:/opt/cray/pe/lmod/modulefiles/craype-targets/default/craype-network-ofi.lua:/opt/cray/pe/lmod/modulefiles/core/perftools-base/22.12.0.lua:/opt/cray/modulefiles/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta:/opt/cray/pe/lmod/modulefiles/core/cray-pmi/6.1.8.lua:/opt/cray/pe/lmod/modulefiles/core/cray-python/3.9.13.1.lua:/opt/cray/pe/lmod/modulefiles/core/amd/5.3.0.lua:/opt/cray/pe/lmod/modulefiles/core/craype/2.7.19.lua:/opt/cray/pe/lmod/modulefiles/core/cray-dsmml/0.2.2.lua:/opt/cray/pe/lmod/modulefiles/comnet/amd/4.0/ofi/1.0/cray-mpich/8.1.23.lua:/opt/cray/pe/lmod/modulefiles/core/cray-libsci/22.12.1.1.lua:/opt/cray/pe/lmod/modulefiles/core/PrgEnv-amd/8.3.3.lua:/sw/frontier/spack-envs/base/modules/spack/cray-sles15-x86_64/cray-mpich/8.1.23-j56azw5/cce/15.0.0/darshan-runtime/3.4.0.lua:/sw/frontier/modulefiles/hsi/default.lua:/sw/frontier/modulefiles/DefApps/default.lua +OLCF_FAMILY_COMPILER=amd +SLURM_SPANK_GPU_SRANGE_SRC=default +MODULESHOME=/opt/cray/pe/lmod/lmod +PKG_CONFIG_PATH=/sw/frontier/spack-envs/base/opt/cray-sles15-zen3/cce-15.0.0/darshan-runtime-3.4.0-t6el25xrwgfg5j65rdrhrs3qjp4ojssp/lib/pkgconfig:/opt/cray/pe/dsmml/0.2.2/dsmml/lib/pkgconfig:/opt/cray/pe/craype/2.7.19/pkg-config:/opt/cray/pe/pmi/6.1.8/lib/pkgconfig:/opt/cray/libfabric/1.15.2.0/lib64/pkgconfig +LMOD_SETTARG_FULL_SUPPORT=no +LIBSCI_BASE_DIR=/opt/cray/pe/libsci/22.12.1.1 +G_BROKEN_FILENAMES=1 +_ModuleTable011_=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 +PE_MPICH_GTL_LIBS_amd_gfx90a=-lmpi_gtl_hsa +HISTSIZE=1000 +CRAYPE_DIR=/opt/cray/pe/craype/2.7.19 +LMOD_PKG=/opt/cray/pe/lmod/lmod +__LMOD_REF_COUNT_PE_MPICH_PKGCONFIG_LIBS=mpich:1 +OFFLOAD_INIT=on_start +__LMOD_REF_COUNT_PE_PKGCONFIG_LIBS=darshan-runtime:1;libsci_mpi:1;libsci:1;mpich:1;dsmml:1 +CPU=x86_64 +CRAYPE_VERSION=2.7.19 +OLCF_FAMILY_CRAYPE_NETWORK=craype-network-ofi +PE_MPICH_GENCOMPILERS_AMD=5.0 +SLURM_SPANK_GPU_POWER_CAP=560 +_ModuleTable005_=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 +LMOD_CMD=/opt/cray/pe/lmod/lmod/libexec/lmod +CRAY_PMI_POST_LINK_OPTS=-L/opt/cray/pe/pmi/6.1.8/lib +WORLDWORK=/lustre/orion/world-shared +SLURM_LOCALID=0 +CRAY_MPICH_PREFIX=/opt/cray/pe/mpich/8.1.23/ofi/amd/5.0 +LESSOPEN=lessopen.sh %s +CRAYPAT_OPTS_EXECUTABLE=libexec64/opts +CRAY_XPMEM_INCLUDE_OPTS=-I/opt/cray/xpmem/2.6.2-2.5_2.22__gd067c3f.shasta/include +PE_LIBSCI_GENCOMPS_GNU_x86_64=91 +LMOD_DIR=/opt/cray/pe/lmod/lmod/libexec +LMOD_FAMILY_MPI=cray-mpich +BASH_FUNC_module%%=() { if [ -z "${LMOD_SH_DBG_ON+x}" ]; then + case "$-" in + *v*x*) + __lmod_sh_dbg='vx' + ;; + *v*) + __lmod_sh_dbg='v' + ;; + *x*) + __lmod_sh_dbg='x' + ;; + esac; + fi; + if [ -n "${__lmod_sh_dbg:-}" ]; then + set +$__lmod_sh_dbg; + echo "Shell debugging temporarily silenced: export LMOD_SH_DBG_ON=1 for Lmod's output" 1>&2; + fi; + eval "$($LMOD_CMD $LMOD_SHELL_PRGM "$@")" && eval "$(${LMOD_SETTARG_CMD:-:} -s sh)"; + __lmod_my_status=$?; + if [ -n "${__lmod_sh_dbg:-}" ]; then + echo "Shell debugging restarted" 1>&2; + set -$__lmod_sh_dbg; + fi; + unset __lmod_sh_dbg; + return $__lmod_my_status +} +BASH_FUNC_ml%%=() { eval "$($LMOD_DIR/ml_cmd "$@")" +} +_=/usr/bin/env diff --git a/results/figs.py b/results/figs.py index b94672b..621a0dc 100644 --- a/results/figs.py +++ b/results/figs.py @@ -96,14 +96,14 @@ for pattern in [ "implicit_managed_GPURdHost_fine", "implicit_managed_GPUWrGPU_coarse", "implicit_managed_GPUWrGPU_fine", - "implicit_managed_HostWrGPU_coarse", + # "implicit_managed_HostWrGPU_coarse", # takes too long "implicit_managed_HostWrGPU_fine", "implicit_managed_GPUWrHost_coarse", "implicit_managed_GPUWrHost_fine", "implicit_mapped_GPURdHost", "implicit_mapped_GPUWrGPU", "implicit_mapped_GPUWrHost", - "implicit_mapped_HostWrGPU", + # "implicit_mapped_HostWrGPU", # segfault "prefetch_managed_GPUToGPU", "prefetch_managed_GPUToHost", "prefetch_managed_HostToGPU", @@ -112,10 +112,15 @@ for pattern in [ for name, (x, t, terr, bw, bwerr) in series.items(): if pattern not in name: continue + fields = name.split("/") + name = f'{fields[1]} -> {fields[2]}' plt.errorbar(x, bw, yerr=bwerr, label=name) output_path = f"{pattern}.pdf" print(f"write {output_path}") plt.xscale('log') + plt.xlabel('transfer size [B]') + plt.ylabel('bandwidth [GB/s]') + plt.title(pattern) lgd = plt.legend(bbox_to_anchor=(1.04, 1)) plt.tight_layout() plt.savefig(output_path, bbox_extra_artists=(lgd,), bbox_inches='tight') \ No newline at end of file diff --git a/results/hipManaged_HostToGPUWriteDst.json b/results/hipManaged_HostToGPUWriteDst.json new file mode 100644 index 0000000..fd759b5 --- /dev/null +++ b/results/hipManaged_HostToGPUWriteDst.json @@ -0,0 +1,3194 @@ +{ + "context": { + "date": "2023-10-09T18:20:54-04:00", + "host_name": "frontier08444", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1817, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [2.57,2.64,4.15], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8629556387495347e+06, + "cpu_time": 2.0656179550264548e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1986567553209928e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8665628466656909e+06, + "cpu_time": 2.0663696851851849e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1944077625443120e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8713425807912019e+06, + "cpu_time": 2.0706635291005303e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1888028638071255e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8663622094204319e+06, + "cpu_time": 2.0654141851851849e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1946436652679252e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8709447342006518e+06, + "cpu_time": 2.0700148227513211e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1892683012627778e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8676336019655024e+06, + "cpu_time": 2.0676160354497354e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1931558696406269e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8665628466656909e+06, + "cpu_time": 2.0663696851851847e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1944077625443120e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5130619254688922e+03, + "cpu_time": 2.5216689420010753e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1262077809843104e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8810230881323494e-03, + "cpu_time": 1.2196021402264745e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8814019733401053e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8691567384901829e+06, + "cpu_time": 2.0703159973544993e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3827250178158488e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8652921753528062e+06, + "cpu_time": 2.0662920529100546e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3918052668883055e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8609479859978869e+06, + "cpu_time": 2.0665982830687833e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.4020574791117795e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8657145948733189e+06, + "cpu_time": 2.0667645740740735e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3908109110098016e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8724842035255048e+06, + "cpu_time": 2.0737255899470886e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3749367736059614e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8667191396479402e+06, + "cpu_time": 2.0687392994708996e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3884670896863397e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8657145948733184e+06, + "cpu_time": 2.0667645740740735e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3908109110098016e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3453826466637347e+03, + "cpu_time": 3.2334801793449142e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0214896631077270e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3278181245215417e-03, + "cpu_time": 1.5630196517134407e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3276685052701098e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8728744742080506e+06, + "cpu_time": 2.0750226534391525e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7480502434248906e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8680769572956562e+06, + "cpu_time": 2.0701003280423265e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7705166192502547e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8731343694699239e+06, + "cpu_time": 2.0753980079365058e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7468364614101276e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8676520065310849e+06, + "cpu_time": 2.0699288968253937e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7725121932276338e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8673209947890851e+06, + "cpu_time": 2.0744366058201077e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7740672577028368e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8698117604587604e+06, + "cpu_time": 2.0729772984126969e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7623965550031494e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8680769572956562e+06, + "cpu_time": 2.0744366058201075e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7705166192502547e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9282221061476753e+03, + "cpu_time": 2.7268388585636258e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3714955796501861e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5660518176595665e-03, + "cpu_time": 1.3154214764684582e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5652060153190511e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 377, + "real_time": 1.8641717095219526e+06, + "cpu_time": 2.0725986949602163e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7577779896897484e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 377, + "real_time": 1.8686068632818428e+06, + "cpu_time": 2.0724383395225464e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7536058891729325e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 377, + "real_time": 1.8733806414214710e+06, + "cpu_time": 2.0770924244031757e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7491373229487691e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 377, + "real_time": 1.8685301762132593e+06, + "cpu_time": 2.0726734270557035e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7536778595894679e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 377, + "real_time": 1.8737839506855679e+06, + "cpu_time": 2.0777470742705581e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7487608423591767e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8696946682248190e+06, + "cpu_time": 2.0745099920424402e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7525919807520188e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8686068632818430e+06, + "cpu_time": 2.0726734270557039e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7536058891729325e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9796690666856143e+03, + "cpu_time": 2.6676501785135069e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7320943772398095e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1285128177982717e-03, + "cpu_time": 1.2859182114071648e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1294713305936770e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 377, + "real_time": 1.8707741517149005e+06, + "cpu_time": 2.0787002679045114e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5031486799154505e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 377, + "real_time": 1.8658408934814949e+06, + "cpu_time": 2.0789032466843540e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5124109579201892e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 377, + "real_time": 1.8671702081777605e+06, + "cpu_time": 2.0804294297082266e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5099103291691326e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 377, + "real_time": 1.8712356428363156e+06, + "cpu_time": 2.0801671591511904e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5022847203072809e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 377, + "real_time": 1.8780788532215843e+06, + "cpu_time": 2.0867325437665717e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.4895233438991159e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8706199498864114e+06, + "cpu_time": 2.0809865294429709e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5034556062422343e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8707741517149005e+06, + "cpu_time": 2.0801671591511904e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5031486799154505e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7638795365021833e+03, + "cpu_time": 3.3002070870237599e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9083511854747849e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5466848767391036e-03, + "cpu_time": 1.5858858480487831e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5427327149807324e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 376, + "real_time": 1.8710608946616475e+06, + "cpu_time": 2.0914761728723464e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0052236340336934e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 376, + "real_time": 1.8744419295662157e+06, + "cpu_time": 2.0908780558510642e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9925879235070646e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 376, + "real_time": 1.8792347209261821e+06, + "cpu_time": 2.0953002845744735e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9747540602805108e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 376, + "real_time": 1.8708218504662968e+06, + "cpu_time": 2.0913191489361664e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0061187262341782e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 376, + "real_time": 1.8706248702073933e+06, + "cpu_time": 2.0909965505319117e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0068564834951773e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8732368531655471e+06, + "cpu_time": 2.0919940425531927e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9971081655101240e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8710608946616475e+06, + "cpu_time": 2.0913191489361662e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0052236340336934e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7018952992733698e+03, + "cpu_time": 1.8638215682522557e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3804160114798727e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9762024716829633e-03, + "cpu_time": 8.9093062902681030e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9728378907791738e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 375, + "real_time": 1.8819343545086060e+06, + "cpu_time": 2.1138424959999989e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3929497560420948e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 375, + "real_time": 1.8815119534653302e+06, + "cpu_time": 2.1132233626666637e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3932624744540611e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 375, + "real_time": 1.8840809645674501e+06, + "cpu_time": 2.1160349253333285e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3913627117409116e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 375, + "real_time": 1.8787842950550839e+06, + "cpu_time": 2.1104088026666734e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3952852421108523e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 375, + "real_time": 1.8756178764936826e+06, + "cpu_time": 2.1101236053333329e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3976407630004957e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8803858888180307e+06, + "cpu_time": 2.1127266383999996e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3941001894696832e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8815119534653302e+06, + "cpu_time": 2.1132233626666637e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3932624744540611e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2641383428849940e+03, + "cpu_time": 2.4791520347424716e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4217276128830013e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7358874911238349e-03, + "cpu_time": 1.1734372018047596e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7371259477442782e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 374, + "real_time": 1.8946148493584932e+06, + "cpu_time": 2.1578309598930469e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7672537253548986e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 374, + "real_time": 1.8944441549259580e+06, + "cpu_time": 2.1576519866310060e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7675030622398639e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 374, + "real_time": 1.8975774867753971e+06, + "cpu_time": 2.1607162887700410e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7629332854856759e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 374, + "real_time": 1.8994590224116731e+06, + "cpu_time": 2.1606327941176319e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7601964233707494e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 374, + "real_time": 1.9013362704467447e+06, + "cpu_time": 2.1624319411764555e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7574711961751586e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8974863567836534e+06, + "cpu_time": 2.1598527941176360e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7630715385252690e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8975774867753971e+06, + "cpu_time": 2.1606327941176323e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7629332854856759e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0092397963728104e+03, + "cpu_time": 2.0577530093535270e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3811620780336350e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5859085287304210e-03, + "cpu_time": 9.5272835952423332e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5856129734418631e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 371, + "real_time": 1.9087553266850060e+06, + "cpu_time": 2.2355748840970341e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4935066078955972e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 371, + "real_time": 1.9003606476379193e+06, + "cpu_time": 2.2310645660377280e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5177736989204800e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 371, + "real_time": 1.9088407443093283e+06, + "cpu_time": 2.2356895202156371e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4932607821057594e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 371, + "real_time": 1.9049925057812422e+06, + "cpu_time": 2.2335786603773665e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5043576120000350e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 371, + "real_time": 1.9191835339826271e+06, + "cpu_time": 2.2439340862533720e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4636567135610485e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9084265516792252e+06, + "cpu_time": 2.2359683433962283e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4945110828965843e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9087553266850063e+06, + "cpu_time": 2.2355748840970346e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4935066078955972e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9434104739934146e+03, + "cpu_time": 4.8333019034923009e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9954120932779775e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6382906472789880e-03, + "cpu_time": 2.1616146390297125e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6316463160650150e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 362, + "real_time": 1.9441467691508755e+06, + "cpu_time": 2.4523824723756928e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0787004527008786e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 362, + "real_time": 1.9455395188074540e+06, + "cpu_time": 2.4528625966850729e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0779282454696572e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 362, + "real_time": 1.9431015896580042e+06, + "cpu_time": 2.4514163232044280e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0792806774292791e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 362, + "real_time": 1.9445438489500557e+06, + "cpu_time": 2.4530929723756835e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0784801798798952e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 362, + "real_time": 1.9430755632762276e+06, + "cpu_time": 2.4520536215469586e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0792951337744083e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9440814579685233e+06, + "cpu_time": 2.4523615972375674e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0787369378508239e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9441467691508755e+06, + "cpu_time": 2.4523824723756923e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0787004527008786e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0387466501227243e+03, + "cpu_time": 6.6609643268094010e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7629736379754508e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3431230767879824e-04, + "cpu_time": 2.7161428128350088e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3423345727430724e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 355, + "real_time": 1.9782959522915557e+06, + "cpu_time": 2.7920240845070556e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.1201600271898322e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 355, + "real_time": 1.9796548268600353e+06, + "cpu_time": 2.7921684140845030e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.1187047070486817e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 355, + "real_time": 1.9781178834152892e+06, + "cpu_time": 2.7924267830985789e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.1203508826068487e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 355, + "real_time": 1.9859714636867736e+06, + "cpu_time": 2.7958366450704173e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.1119658951260357e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 355, + "real_time": 1.9828273257253771e+06, + "cpu_time": 2.7930298028168846e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.1153148060765197e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9809734903958063e+06, + "cpu_time": 2.7930971459154878e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.1172992636095839e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9796548268600355e+06, + "cpu_time": 2.7924267830985789e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.1187047070486817e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3719066376684068e+03, + "cpu_time": 1.5789854825728460e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6005037531101117e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7021462700112589e-03, + "cpu_time": 5.6531706563872663e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7005171706204404e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 343, + "real_time": 2.0311437482720255e+06, + "cpu_time": 3.5237723790087365e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.1299922800326271e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 343, + "real_time": 2.0381030609624858e+06, + "cpu_time": 3.5271094227405270e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.1158899962784586e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 343, + "real_time": 2.0356778827232495e+06, + "cpu_time": 3.5243348921282855e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.1207934080307689e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 343, + "real_time": 2.0409667207357141e+06, + "cpu_time": 3.5331205772594619e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.1101150326332273e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 343, + "real_time": 2.0338897810537510e+06, + "cpu_time": 3.5246654810495563e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.1244162186870780e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0359562387494452e+06, + "cpu_time": 3.5266005504373135e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.1202413871324320e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0356778827232495e+06, + "cpu_time": 3.5246654810495563e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.1207934080307689e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7822469984866775e+03, + "cpu_time": 3.8615397028022426e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6532920325203845e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8577250957072947e-03, + "cpu_time": 1.0949750751678964e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8574863250540893e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 194, + "real_time": 3.5845442703857864e+06, + "cpu_time": 6.5146454793814430e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.6804320813129063e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 194, + "real_time": 3.5879430553117362e+06, + "cpu_time": 6.5200649175257869e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.6759984039218035e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 194, + "real_time": 3.5988585699407249e+06, + "cpu_time": 6.5327575824742056e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.6618158713239822e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 194, + "real_time": 3.5820616476391396e+06, + "cpu_time": 6.5061893092783308e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.6836759526619272e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 194, + "real_time": 3.5824319347739220e+06, + "cpu_time": 6.5055173298969222e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.6831918388028679e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5871678956102617e+06, + "cpu_time": 6.5158349237113371e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.6770228296046972e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5845442703857860e+06, + "cpu_time": 6.5146454793814421e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.6804320813129063e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9400872032627040e+03, + "cpu_time": 1.1238045037008063e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.0311232172482293e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9346981811906605e-03, + "cpu_time": 1.7247283224000119e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9309555557614289e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 98, + "real_time": 7.2482181855534408e+06, + "cpu_time": 1.3398360602040812e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.6293352574399548e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 98, + "real_time": 7.2532190266540460e+06, + "cpu_time": 1.3390922877550997e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.6261434925230513e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 98, + "real_time": 7.2480548572327411e+06, + "cpu_time": 1.3394521561224483e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.6294395752974281e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 98, + "real_time": 7.2535649464674750e+06, + "cpu_time": 1.3363635795918457e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.6259228734611645e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 98, + "real_time": 7.2538033222817648e+06, + "cpu_time": 1.3405745061224446e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.6257708555358353e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2513720676378934e+06, + "cpu_time": 1.3390637179591838e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.6273224108514862e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2532190266540470e+06, + "cpu_time": 1.3394521561224481e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.6261434925230513e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9614958783472366e+03, + "cpu_time": 1.6063499821469439e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8900935131998099e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.0840489920026027e-04, + "cpu_time": 1.1996068302075429e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0846376054700035e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50, + "real_time": 1.4162119254469872e+07, + "cpu_time": 2.5764457139999878e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7386173491526690e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50, + "real_time": 1.4169204011559486e+07, + "cpu_time": 2.5764054400000021e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7362479886132917e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50, + "real_time": 1.4161895010620356e+07, + "cpu_time": 2.5747192400000077e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7386923818933411e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50, + "real_time": 1.4160922113806009e+07, + "cpu_time": 2.5743086479999989e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7390179439355202e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50, + "real_time": 1.4158444982022047e+07, + "cpu_time": 2.5726607599999908e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7398470725572443e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4162517074495558e+07, + "cpu_time": 2.5749079603999976e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7384845472304134e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4161895010620356e+07, + "cpu_time": 2.5747192400000073e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7386923818933411e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0120502289617848e+03, + "cpu_time": 1.5852311869884707e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3420146574460354e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8328652370607552e-04, + "cpu_time": 6.1564576729267401e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8321600378129893e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 26, + "real_time": 2.7426020958675787e+07, + "cpu_time": 4.8683463807691939e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.8938097218780966e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 26, + "real_time": 2.7459080187747113e+07, + "cpu_time": 4.8771551653846227e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.8879178429250927e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 26, + "real_time": 2.7422913636725683e+07, + "cpu_time": 4.8779078884615175e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.8943642450979805e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 26, + "real_time": 2.7457086297755059e+07, + "cpu_time": 4.8794694307692163e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.8882727957545109e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 26, + "real_time": 2.7354598905031499e+07, + "cpu_time": 4.8598062923076667e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.9065873152068958e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7423939997187030e+07, + "cpu_time": 4.8725370315384433e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.8941903841725149e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7426020958675791e+07, + "cpu_time": 4.8771551653846227e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.8938097218780966e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2270205560003029e+04, + "cpu_time": 8.3346370442866610e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5535034401792660e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5413615098464636e-03, + "cpu_time": 1.7105333402987193e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5433611787164618e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10, + "real_time": 5.3539996221661568e+07, + "cpu_time": 9.1677904999998823e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0137369246095428e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10, + "real_time": 5.3630397096276283e+07, + "cpu_time": 9.1670751600000232e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0052856315441723e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10, + "real_time": 5.3598203137516975e+07, + "cpu_time": 9.1685681699999586e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0082920748532343e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10, + "real_time": 5.3412138298153877e+07, + "cpu_time": 9.1585909600000545e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0257388030705023e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10, + "real_time": 5.3483483195304871e+07, + "cpu_time": 9.1713879499999970e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0190346619676600e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3532843589782722e+07, + "cpu_time": 9.1666825479999840e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0144176192090225e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3539996221661568e+07, + "cpu_time": 9.1677904999998823e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0137369246095428e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7815051701797041e+04, + "cpu_time": 4.8101942888259822e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2289902988689933e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6403957984133209e-03, + "cpu_time": 5.2474755874201034e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6410660068171665e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 1.0493292007595301e+08, + "cpu_time": 1.7700198500000042e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1163249017696238e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 1.0505743976682425e+08, + "cpu_time": 1.7691265624999985e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1102607601288300e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 1.0514360293745995e+08, + "cpu_time": 1.7705691799999991e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1060729992231112e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 1.0501736216247082e+08, + "cpu_time": 1.7725172799999988e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1122109805939980e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 1.0524440463632345e+08, + "cpu_time": 1.7728941512500107e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1011824700342073e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0507914591580629e+08, + "cpu_time": 1.7710254047500020e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1092104223499537e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0505743976682425e+08, + "cpu_time": 1.7705691799999991e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1102607601288300e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1951520577584048e+05, + "cpu_time": 1.6234830232242894e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8097723475399623e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1373827293153006e-03, + "cpu_time": 9.1669098527328012e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1371174540248802e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.8476798012852669e+08, + "cpu_time": 3.0631960449999696e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.8112981656945810e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.8438761308789253e+08, + "cpu_time": 3.0591839700000280e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.8232860983355570e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.8369403854012489e+08, + "cpu_time": 3.0547271400000042e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.8452731102945347e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.8437624350190163e+08, + "cpu_time": 3.0619248849999535e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.8236451920603619e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.8343342840671539e+08, + "cpu_time": 3.0506529249999660e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.8535776893362093e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8413186073303223e+08, + "cpu_time": 3.0579369929999840e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.8314160511442490e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8437624350190163e+08, + "cpu_time": 3.0591839700000274e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.8236451920603619e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4983544745712623e+05, + "cpu_time": 5.2076065780742478e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7423891601646286e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipManaged_HostToGPUWriteDst/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9860961881785228e-03, + "cpu_time": 1.7029803393579189e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9879349113200978e-03, + "dst_gpu": NaN, + "src_numa": NaN + } + ] +} diff --git a/results/hipMemcpyAsync_GPUToGPU.json b/results/hipMemcpyAsync_GPUToGPU.json new file mode 100644 index 0000000..b937a98 --- /dev/null +++ b/results/hipMemcpyAsync_GPUToGPU.json @@ -0,0 +1,30584 @@ +{ + "context": { + "date": "2023-10-09T17:38:31-04:00", + "host_name": "frontier08319", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1961, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.68,11.22,26.93], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 100306, + "real_time": 7.0115232196496581e+03, + "cpu_time": 1.7180502973939946e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.6511324569612056e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 100306, + "real_time": 7.0013590012837394e+03, + "cpu_time": 1.7174675964965777e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.6564329861254208e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 100306, + "real_time": 7.0336223678249053e+03, + "cpu_time": 1.7189156118925445e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.6396608548543110e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 100306, + "real_time": 7.0159969611340994e+03, + "cpu_time": 1.7170952740643780e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.6488043170220949e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 100306, + "real_time": 7.0274092556332917e+03, + "cpu_time": 1.7187533739963179e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.6428787720707461e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0179821611051393e+03, + "cpu_time": 1.7180564307687631e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.6477818774067558e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0159969611341003e+03, + "cpu_time": 1.7180502973939948e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.6488043170220949e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2801118736088958e+01, + "cpu_time": 7.8970052727578883e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6540381987181288e+04, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8240454937367828e-03, + "cpu_time": 4.5964760710591370e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8241326982655418e-03, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 99014, + "real_time": 7.0852197512046823e+03, + "cpu_time": 1.6989001102301784e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.2263107987997964e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 99014, + "real_time": 7.0739946192906482e+03, + "cpu_time": 1.7004376231195815e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.2377776285521299e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 99014, + "real_time": 7.0779500823026092e+03, + "cpu_time": 1.7002414857177010e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.2337328470312610e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 99014, + "real_time": 7.0675296052504127e+03, + "cpu_time": 1.6963044902382383e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.2443983767629236e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 99014, + "real_time": 7.0854696352414139e+03, + "cpu_time": 1.6999817366956122e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.2260559477022618e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0780327386579538e+03, + "cpu_time": 1.6991730892002621e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.2336551197696760e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0779500823026083e+03, + "cpu_time": 1.6999817366956124e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.2337328470312610e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6418196712163908e+00, + "cpu_time": 1.7101027340209875e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8119324993883551e+04, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0796530552167711e-03, + "cpu_time": 1.0064323316383674e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0799426251381877e-03, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 97132, + "real_time": 7.2500895994442126e+03, + "cpu_time": 1.6753079546751318e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4123963379411191e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 97132, + "real_time": 7.1860972608262355e+03, + "cpu_time": 1.6731206365759637e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4249737553402716e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 97132, + "real_time": 7.2847460179230111e+03, + "cpu_time": 1.6796620460613875e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4056770098512748e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 97132, + "real_time": 7.1905515783159053e+03, + "cpu_time": 1.6718684420226898e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4240910295227039e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 97132, + "real_time": 7.2421589173558932e+03, + "cpu_time": 1.6747466466349235e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4139430129681021e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2307286747730523e+03, + "cpu_time": 1.6749411451940193e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4162162291246945e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2421589173558923e+03, + "cpu_time": 1.6747466466349235e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4139430129681021e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1921150393357088e+01, + "cpu_time": 2.9671130872822320e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2088514721001009e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.7976384233049448e-03, + "cpu_time": 1.7714730429758069e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7963263683072303e-03, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 87997, + "real_time": 8.0350135442169012e+03, + "cpu_time": 1.5524673304284048e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.5488444900930154e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 87997, + "real_time": 7.9274158857984030e+03, + "cpu_time": 1.5489519855426276e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.5834395842268059e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 87997, + "real_time": 8.1331332027676508e+03, + "cpu_time": 1.5588641858580942e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.5180947476712656e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 87997, + "real_time": 7.9961045002776964e+03, + "cpu_time": 1.5500143511326079e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.5612471672035742e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 87997, + "real_time": 8.0791131948589327e+03, + "cpu_time": 1.5568050139355564e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.5349316819861189e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0341560655839157e+03, + "cpu_time": 1.5534205733794582e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.5493115342361560e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0350135442169003e+03, + "cpu_time": 1.5524673304284043e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.5488444900930154e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8517100096481101e+01, + "cpu_time": 4.2885502815381559e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4943281534794094e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.7729119842252573e-03, + "cpu_time": 2.7607142296361105e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.7843206684693363e-03, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 73193, + "real_time": 9.6062291764266374e+03, + "cpu_time": 1.3498673660685866e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.2638999390639621e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 73193, + "real_time": 9.5327481703440098e+03, + "cpu_time": 1.3473130162010248e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.2967672352265513e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 73193, + "real_time": 9.7229487389374626e+03, + "cpu_time": 1.3554789862597060e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.2127137661404729e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 73193, + "real_time": 9.5508084222488451e+03, + "cpu_time": 1.3467703039974911e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.2886421954169518e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 73193, + "real_time": 9.6261709722520045e+03, + "cpu_time": 1.3515101209977055e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.2550667464840972e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6077810960417937e+03, + "cpu_time": 1.3501879587049026e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.2634179764664072e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6062291764266374e+03, + "cpu_time": 1.3498673660685868e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.2638999390639621e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4966671710602270e+01, + "cpu_time": 3.5281554476972716e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3125996424439824e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.8027039710018972e-03, + "cpu_time": 2.6130846634726838e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.7698214454439227e-03, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 65037, + "real_time": 1.0776668717211567e+04, + "cpu_time": 1.2357186778778684e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.6016069668323779e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 65037, + "real_time": 1.0765819969326825e+04, + "cpu_time": 1.2355149771362653e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.6092671281333315e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 65037, + "real_time": 1.0777033416965553e+04, + "cpu_time": 1.2364431985786688e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.6013497249659538e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 65037, + "real_time": 1.0764781130401898e+04, + "cpu_time": 1.2341070476690447e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.6100014489511096e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 65037, + "real_time": 1.0774784482914896e+04, + "cpu_time": 1.2362193247694230e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.6029362935190904e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0771817543364148e+04, + "cpu_time": 1.2356006452062542e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.6050323124803734e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0774784482914898e+04, + "cpu_time": 1.2357186778778684e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.6029362935190904e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.0212946493342177e+00, + "cpu_time": 9.1443492388734664e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2518737027338898e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.5898594875881133e-04, + "cpu_time": 7.4007320037834993e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5908686880347337e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 64812, + "real_time": 1.0816428586753922e+04, + "cpu_time": 1.2498927933422191e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.5147328777323291e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 64812, + "real_time": 1.0810178249540104e+04, + "cpu_time": 1.2493083417537525e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.5156086811701763e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 64812, + "real_time": 1.0822037546611693e+04, + "cpu_time": 1.2501903234149446e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.5139478059868419e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 64812, + "real_time": 1.0803228826751185e+04, + "cpu_time": 1.2469105080027103e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.5165836309446294e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 64812, + "real_time": 1.0815204704942293e+04, + "cpu_time": 1.2487039607915609e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.5149042895612414e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0813415582919841e+04, + "cpu_time": 1.2490011854610379e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.5151554570790436e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0815204704942293e+04, + "cpu_time": 1.2493083417537527e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.5149042895612414e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0863907623505851e+00, + "cpu_time": 1.3002531880643517e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9312288464217761e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.5533324859388378e-04, + "cpu_time": 1.0410343906794577e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5545939857336219e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 64610, + "real_time": 1.0849094845533495e+04, + "cpu_time": 1.2667681982162707e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0203441362198415e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 64610, + "real_time": 1.0831473013372368e+04, + "cpu_time": 1.2676589453952892e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0252579644103007e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 64610, + "real_time": 1.0845674982179344e+04, + "cpu_time": 1.2681220863039889e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0212965125583687e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 64610, + "real_time": 1.0832048823941672e+04, + "cpu_time": 1.2672441877700455e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0250971476027803e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 64610, + "real_time": 1.0847805754404446e+04, + "cpu_time": 1.2678119381262810e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0207030566246519e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0841219483886267e+04, + "cpu_time": 1.2675210711623755e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0225397634831886e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0845674982179346e+04, + "cpu_time": 1.2676589453952892e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0212965125583687e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7227638504145446e+00, + "cpu_time": 5.2638250857141160e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4325282034295099e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.0459249657103009e-04, + "cpu_time": 4.1528501619992366e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.0479609658675041e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 61539, + "real_time": 1.1383167277143144e+04, + "cpu_time": 1.2625428742786050e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.7572728577566595e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 61539, + "real_time": 1.1377577474836349e+04, + "cpu_time": 1.2623261453715994e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.7601014051493101e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 61539, + "real_time": 1.1386309593855718e+04, + "cpu_time": 1.2626978936962218e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.7556840045315952e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 61539, + "real_time": 1.1375098750975887e+04, + "cpu_time": 1.2601894258494501e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.7613565767398338e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 61539, + "real_time": 1.1386314224521117e+04, + "cpu_time": 1.2626223967059653e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.7556816637700253e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1381693464266446e+04, + "cpu_time": 1.2620757471803687e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.7580193015894852e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1383167277143144e+04, + "cpu_time": 1.2625428742786052e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.7572728577566595e+09, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1299077583849408e+00, + "cpu_time": 1.0636032077708056e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5955684335266524e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.5071568431276192e-04, + "cpu_time": 8.4274118264852579e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5077452811076077e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54375, + "real_time": 1.2893414342054701e+04, + "cpu_time": 1.2261151488000429e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0165809964896568e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54375, + "real_time": 1.2880211020987108e+04, + "cpu_time": 1.2259497520491643e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0176230792060034e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54375, + "real_time": 1.2881667749309274e+04, + "cpu_time": 1.2264187621497712e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0175080009110479e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54375, + "real_time": 1.2876933633253491e+04, + "cpu_time": 1.2244167916496956e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0178820807270348e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54375, + "real_time": 1.2882748025823412e+04, + "cpu_time": 1.2252878925946929e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0174226782769232e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2882994954285597e+04, + "cpu_time": 1.2256376694486732e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0174033671221334e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2881667749309276e+04, + "cpu_time": 1.2259497520491643e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0175080009110479e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2224590297675739e+00, + "cpu_time": 7.9818142273215717e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9118599359835172e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.8299786282984141e-04, + "cpu_time": 6.5123767213454023e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8278392766454021e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53698, + "real_time": 1.3046322704092983e+04, + "cpu_time": 1.2979320229976521e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0093324835339104e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53698, + "real_time": 1.3038168084236833e+04, + "cpu_time": 1.2974965904916525e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0105892047590069e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53698, + "real_time": 1.3046573491506813e+04, + "cpu_time": 1.2986794555146058e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0092938591933971e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53698, + "real_time": 1.3031595051497656e+04, + "cpu_time": 1.2961640085114226e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0116033299382881e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53698, + "real_time": 1.3044990993523223e+04, + "cpu_time": 1.2975662067683339e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0095376081911690e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3041530064971503e+04, + "cpu_time": 1.2975676568567336e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0100712971231544e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3044990993523224e+04, + "cpu_time": 1.2975662067683339e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0095376081911690e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.5266946962057215e+00, + "cpu_time": 9.1434951550149336e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0062842532612741e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.0045467546295792e-04, + "cpu_time": 7.0466423131757211e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0062117433420596e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53104, + "real_time": 1.3198719934167060e+04, + "cpu_time": 1.2901886503573179e+09, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.9722639969258995e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53104, + "real_time": 1.3185339775302487e+04, + "cpu_time": 1.2892527473050745e+09, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.9762949528388031e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53104, + "real_time": 1.3200862300402732e+04, + "cpu_time": 1.2904750789833174e+09, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.9716193387155098e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53104, + "real_time": 1.3193335783163811e+04, + "cpu_time": 1.2893452240863063e+09, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.9738850630107574e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53104, + "real_time": 1.3198877048794486e+04, + "cpu_time": 1.2901110597529433e+09, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.9722167125413567e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3195426968366117e+04, + "cpu_time": 1.2898745520969920e+09, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.9732560128064659e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3198719934167060e+04, + "cpu_time": 1.2901110597529433e+09, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.9722639969258995e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2931318211050051e+00, + "cpu_time": 5.4361187298799690e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8956399505290028e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.7691763488909923e-04, + "cpu_time": 4.2144553678047917e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7709987587485917e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51414, + "real_time": 1.3631755214313680e+04, + "cpu_time": 1.2661333444642556e+09, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6921569050694901e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51414, + "real_time": 1.3617996876640656e+04, + "cpu_time": 1.2646225310570486e+09, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6999283337966751e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51414, + "real_time": 1.3632281094236872e+04, + "cpu_time": 1.2650260500399606e+09, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6918601718335434e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51414, + "real_time": 1.3619496466909070e+04, + "cpu_time": 1.2626089097861156e+09, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6990805243622437e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51414, + "real_time": 1.3628872721802036e+04, + "cpu_time": 1.2647686454652176e+09, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6937837883143387e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3626080474780465e+04, + "cpu_time": 1.2646318961625197e+09, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6953619446752579e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3628872721802036e+04, + "cpu_time": 1.2647686454652176e+09, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6937837883143387e+10, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8399691592809004e+00, + "cpu_time": 1.2768091319886786e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8634509811336286e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.0197627791355762e-04, + "cpu_time": 1.0096290753563233e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0204928746813682e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 47790, + "real_time": 1.4651240174334325e+04, + "cpu_time": 1.2083656046587665e+09, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4313819001300232e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 47790, + "real_time": 1.4655323838088410e+04, + "cpu_time": 1.2077182151058910e+09, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4309830496884778e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 47790, + "real_time": 1.4674059596266075e+04, + "cpu_time": 1.2087914624697897e+09, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4291559784407828e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 47790, + "real_time": 1.4657824309321833e+04, + "cpu_time": 1.2076360409030101e+09, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4307389389749265e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 47790, + "real_time": 1.4681114308741326e+04, + "cpu_time": 1.2094978405854466e+09, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4284692264478375e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4663912445350396e+04, + "cpu_time": 1.2084018327445807e+09, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4301458187364096e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4657824309321830e+04, + "cpu_time": 1.2083656046587665e+09, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4307389389749265e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2944929543780308e+01, + "cpu_time": 7.7590708307452861e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2620806241621174e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.8277460684681478e-04, + "cpu_time": 6.4209360003389848e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8248387515981797e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 37997, + "real_time": 1.8424056943008723e+04, + "cpu_time": 1.0566001625617011e+09, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2765366026463513e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 37997, + "real_time": 1.8412881136884047e+04, + "cpu_time": 1.0556889334256701e+09, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2779183598802011e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 37997, + "real_time": 1.8418903430935490e+04, + "cpu_time": 1.0555633598943887e+09, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2771735655855887e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 37997, + "real_time": 1.8418920747310571e+04, + "cpu_time": 1.0553124781335189e+09, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2771714247222812e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 37997, + "real_time": 1.8420182846863328e+04, + "cpu_time": 1.0565358885182704e+09, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2770153992875400e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8418989021000434e+04, + "cpu_time": 1.0559401645067098e+09, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2771630704243927e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8418920747310574e+04, + "cpu_time": 1.0556889334256703e+09, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2771714247222812e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0133305782516562e+00, + "cpu_time": 5.8940085828104988e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9621334712754354e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1789092624333793e-04, + "cpu_time": 5.5817637977279976e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1790856947063731e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 28043, + "real_time": 2.5003900296529966e+04, + "cpu_time": 9.2663917864539385e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3549197927189661e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 28043, + "real_time": 2.4965906793595623e+04, + "cpu_time": 9.2597498430909276e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3600253615269794e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 28043, + "real_time": 2.4996367552396383e+04, + "cpu_time": 9.2633629747259188e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3559308097131061e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 28043, + "real_time": 2.4943636775950781e+04, + "cpu_time": 9.2513116761005640e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3630252377984485e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 28043, + "real_time": 2.5002330096072928e+04, + "cpu_time": 9.2611941563117528e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3551304889449420e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4982428302909142e+04, + "cpu_time": 9.2604020873366201e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3578063381404883e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4996367552396387e+04, + "cpu_time": 9.2611941563117528e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3559308097131061e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6594987266358856e+01, + "cpu_time": 5.6641795990029140e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5766918619899255e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0645477270623024e-03, + "cpu_time": 6.1165590279805941e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0651870601836595e-03, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18729, + "real_time": 3.7389143687856515e+04, + "cpu_time": 8.4144416498611701e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.4871891531040900e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18729, + "real_time": 3.7381721409837308e+04, + "cpu_time": 8.4141788072469807e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.4880801009835083e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18729, + "real_time": 3.7397489995982061e+04, + "cpu_time": 8.4165619457734013e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.4861877098710431e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18729, + "real_time": 3.7370707629749071e+04, + "cpu_time": 8.4095701433296311e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.4894028141560919e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18729, + "real_time": 3.7390543805830675e+04, + "cpu_time": 8.4152531032961369e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.4870211268186383e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7385921305851130e+04, + "cpu_time": 8.4140011299014640e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.4875761809866742e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7389143687856515e+04, + "cpu_time": 8.4144416498611712e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.4871891531040900e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0181857764883633e+01, + "cpu_time": 2.6450181667074090e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2223186835462248e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7234470648955516e-04, + "cpu_time": 3.1435914089761769e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7237836958067551e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10683, + "real_time": 6.5417162606758837e+04, + "cpu_time": 7.5656835028194702e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1293010370543323e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10683, + "real_time": 6.5290222886376934e+04, + "cpu_time": 7.5659670779453671e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1392736181026685e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10683, + "real_time": 6.5323112176298047e+04, + "cpu_time": 7.5607375070420265e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1366860644118164e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10683, + "real_time": 6.5300882181308662e+04, + "cpu_time": 7.5645009792164528e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1384347162165021e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10683, + "real_time": 6.5303278430736835e+04, + "cpu_time": 7.5607492611101758e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1382461656330341e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.5326931656295863e+04, + "cpu_time": 7.5635276656266999e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1363883202836713e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.5303278430736842e+04, + "cpu_time": 7.5645009792164528e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1382461656330341e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1821320095313851e+01, + "cpu_time": 2.6004914130506851e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0707033868842351e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.9326119842825329e-04, + "cpu_time": 3.4381991155646994e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.9252251446974311e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5892, + "real_time": 1.1885753989272064e+05, + "cpu_time": 7.1589378420061564e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.6461596008609668e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5892, + "real_time": 1.1876594207939696e+05, + "cpu_time": 7.1548775553342247e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.6505141815097656e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5892, + "real_time": 1.1883486376325136e+05, + "cpu_time": 7.1578258898625970e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.6472370039231567e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5892, + "real_time": 1.1876053861269068e+05, + "cpu_time": 7.1533124568380821e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.6507712733486023e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5892, + "real_time": 1.1882492456928684e+05, + "cpu_time": 7.1597259060012066e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.6477093710140588e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1880876178346931e+05, + "cpu_time": 7.1569359300084555e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.6484782861313098e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1882492456928684e+05, + "cpu_time": 7.1578258898625970e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.6477093710140588e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3245965667810069e+01, + "cpu_time": 2.7363346971472545e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0561477166227451e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6399643442649852e-04, + "cpu_time": 3.8233326718407852e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6401799076951359e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3101, + "real_time": 2.2569147678334187e+05, + "cpu_time": 6.9013456732311940e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.9469559911137292e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3101, + "real_time": 2.2567352051770099e+05, + "cpu_time": 6.9047030233925748e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.9474291752129810e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3101, + "real_time": 2.2573114963087995e+05, + "cpu_time": 6.9026157040600514e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.9459107978440491e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3101, + "real_time": 2.2565122918911779e+05, + "cpu_time": 6.9029930609028459e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.9480167018063269e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3101, + "real_time": 2.2566247857041369e+05, + "cpu_time": 6.9019165691100883e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.9477201903603088e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2568197093829085e+05, + "cpu_time": 6.9027148061393523e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.9472065712674792e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2567352051770096e+05, + "cpu_time": 6.9026157040600514e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.9474291752129810e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1246288111626978e+01, + "cpu_time": 1.2798400296012987e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2332978111789927e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3845274384000650e-04, + "cpu_time": 1.8541111222833585e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3843974835103630e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1574, + "real_time": 4.4493275761336269e+05, + "cpu_time": 6.7428286028718770e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 6.0331690891877380e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1574, + "real_time": 4.4466082453124190e+05, + "cpu_time": 6.7392232660229242e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 6.0368586839864429e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1574, + "real_time": 4.4488112063216954e+05, + "cpu_time": 6.7421923511877811e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 6.0338693540997473e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1574, + "real_time": 4.4460706576247007e+05, + "cpu_time": 6.7393669045618439e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 6.0375886186075769e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1574, + "real_time": 4.4502841038457822e+05, + "cpu_time": 6.7435007884372187e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 6.0318723419933423e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4482203578476456e+05, + "cpu_time": 6.7414223826163292e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 6.0346716175749695e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4488112063216948e+05, + "cpu_time": 6.7421923511877799e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 6.0338693540997473e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8065297985729896e+02, + "cpu_time": 1.9969465193473193e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4509853408437416e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.0612416949755450e-04, + "cpu_time": 2.9622035321458516e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0615057391120630e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 792, + "real_time": 8.7271568205470988e+05, + "cpu_time": 6.6087480433077967e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 6.1517275676311719e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 792, + "real_time": 8.7271366837831924e+05, + "cpu_time": 6.6080539050758660e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 6.1517417619643347e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 792, + "real_time": 8.7268497692562186e+05, + "cpu_time": 6.6083143336869168e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 6.1519440141085071e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 792, + "real_time": 8.7266173325431999e+05, + "cpu_time": 6.6079861233837700e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 6.1521078734357617e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 792, + "real_time": 8.7282638773416181e+05, + "cpu_time": 6.6089872450252712e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 6.1509473080174072e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7272048966942669e+05, + "cpu_time": 6.6084179300959253e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 6.1516937050314368e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7271366837831936e+05, + "cpu_time": 6.6083143336869168e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 6.1517417619643347e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3240640952248285e+01, + "cpu_time": 4.3669143626568512e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4574508989162855e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.2463797631476358e-05, + "cpu_time": 6.6081086409035031e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.2458921276762536e-05, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 397, + "real_time": 1.7466940773388264e+06, + "cpu_time": 6.5986659537027633e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.1472803848736816e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 397, + "real_time": 1.7438632467179850e+06, + "cpu_time": 6.5926642286901736e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.1572593265029333e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 397, + "real_time": 1.7466388999951312e+06, + "cpu_time": 6.5985208390430260e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.1474745810538916e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 397, + "real_time": 1.7437854592901859e+06, + "cpu_time": 6.5929295532241571e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.1575339918080884e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 397, + "real_time": 1.7463011638981570e+06, + "cpu_time": 6.5977778174057245e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.1486635077488843e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7454565694480571e+06, + "cpu_time": 6.5961116784131706e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.1516423583974963e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7463011638981570e+06, + "cpu_time": 6.5977778174057257e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.1486635077488843e+11, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4978282250383413e+03, + "cpu_time": 3.0461133678426931e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2804452250178510e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.5812975885843996e-04, + "cpu_time": 4.6180439573386637e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.5837974924039763e-04, + "dst_id": NaN, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56881, + "real_time": 1.2079012820381886e+04, + "cpu_time": 1.1834519169688873e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1193784939778417e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56881, + "real_time": 1.2015181963764493e+04, + "cpu_time": 1.1757437588658333e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1306377279349361e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56881, + "real_time": 1.2158100591296514e+04, + "cpu_time": 1.1771198788612449e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1055920542659428e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56881, + "real_time": 1.2148390505384265e+04, + "cpu_time": 1.1821263120536966e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1072750327423107e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56881, + "real_time": 1.2029115520195128e+04, + "cpu_time": 1.1804303645635259e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1281697691755757e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2085960280204457e+04, + "cpu_time": 1.1797744462626376e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1182106156193215e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time_median", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2079012820381886e+04, + "cpu_time": 1.1804303645635259e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1193784939778417e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.5937321595710785e+01, + "cpu_time": 3.2702622713185856e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1552022154286451e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.4556957053474055e-03, + "cpu_time": 2.7719385528973996e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4536702200922906e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57942, + "real_time": 1.2210898973387768e+04, + "cpu_time": 1.2009049973787222e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1929754812962122e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57942, + "real_time": 1.2241142124460659e+04, + "cpu_time": 1.2062118762695320e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1826162525872849e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57942, + "real_time": 1.2088622113745909e+04, + "cpu_time": 1.2042060620629275e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2353875833194211e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57942, + "real_time": 1.2082968730373850e+04, + "cpu_time": 1.2001710664883387e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2373692378508590e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57942, + "real_time": 1.2200187802627348e+04, + "cpu_time": 1.2000307125681956e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1966567095773660e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2164763948919108e+04, + "cpu_time": 1.2023049429535432e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2090010529262289e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time_median", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2200187802627348e+04, + "cpu_time": 1.2009049973787220e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1966567095773660e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.3663003661414436e+01, + "cpu_time": 2.7642260232708883e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5526272494236208e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.0554404483910860e-03, + "cpu_time": 2.2991055966886239e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0646866496955488e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56993, + "real_time": 1.2141257151324615e+04, + "cpu_time": 1.1889377982051513e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.4340524810339049e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56993, + "real_time": 1.2112372535262140e+04, + "cpu_time": 1.1818185562863734e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.4541653339911759e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56993, + "real_time": 1.2319969968233148e+04, + "cpu_time": 1.1853685291737854e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.3117085726699665e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56993, + "real_time": 1.2265933742916979e+04, + "cpu_time": 1.1898983238159921e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.3483248928465277e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56993, + "real_time": 1.2168750464543797e+04, + "cpu_time": 1.1880755923444893e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.4149971107028484e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2201656772456136e+04, + "cpu_time": 1.1868197599651587e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.3926496782488853e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time_median", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2168750464543798e+04, + "cpu_time": 1.1880755923444893e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.4149971107028484e+07, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7795437767617031e+01, + "cpu_time": 3.2656877293455256e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0236875936588878e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.1953702193791689e-03, + "cpu_time": 2.7516290505993904e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1773371040023222e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57151, + "real_time": 1.2403729385293549e+04, + "cpu_time": 1.1914515660181074e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6511163186357534e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57151, + "real_time": 1.2359646424861961e+04, + "cpu_time": 1.1949731937718859e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6570053297644177e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57151, + "real_time": 1.2264616298272997e+04, + "cpu_time": 1.1953134511866221e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6698443311988348e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57151, + "real_time": 1.2242034684708055e+04, + "cpu_time": 1.1897572252649004e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6729245201030403e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57151, + "real_time": 1.2389476130008872e+04, + "cpu_time": 1.1906321417386580e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6530158164149377e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2331900584629089e+04, + "cpu_time": 1.1924255155960350e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6607812632233968e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time_median", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2359646424861963e+04, + "cpu_time": 1.1914515660181077e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6570053297644177e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.3904003342141792e+01, + "cpu_time": 2.5551603141938173e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9695646880392928e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.9929126767578975e-03, + "cpu_time": 2.1428259298163525e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0029366351890589e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55823, + "real_time": 1.2401944754128041e+04, + "cpu_time": 1.1718298312374570e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3027078262355822e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55823, + "real_time": 1.2378879473216202e+04, + "cpu_time": 1.1648317137641509e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3088616856334925e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55823, + "real_time": 1.2523437820245930e+04, + "cpu_time": 1.1676158638498850e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2706674148038083e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55823, + "real_time": 1.2533642085965679e+04, + "cpu_time": 1.1742114390021338e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2680046006630605e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55823, + "real_time": 1.2402809883025582e+04, + "cpu_time": 1.1716937893870633e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3024774536016738e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2448142803316285e+04, + "cpu_time": 1.1700365274481380e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2905437961875242e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time_median", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2402809883025584e+04, + "cpu_time": 1.1716937893870633e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3024774536016738e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4105008628638586e+01, + "cpu_time": 3.7530667407077639e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9551257963517334e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.9530975664013434e-03, + "cpu_time": 3.2076492080920262e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9416495188940287e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55946, + "real_time": 1.2661895965209909e+04, + "cpu_time": 1.1747272338255675e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4698051717598295e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55946, + "real_time": 1.2673448073836893e+04, + "cpu_time": 1.1793000065780091e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4639078112542963e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55946, + "real_time": 1.2563999975952918e+04, + "cpu_time": 1.1775370233729839e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.5202165040426755e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55946, + "real_time": 1.2554846806935908e+04, + "cpu_time": 1.1727954543821149e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.5249700979818738e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55946, + "real_time": 1.2676373405086222e+04, + "cpu_time": 1.1741199693775041e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4624161329241633e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2626112845404372e+04, + "cpu_time": 1.1756959375072360e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4882631435925674e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time_median", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2661895965209911e+04, + "cpu_time": 1.1747272338255675e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4698051717598295e+08, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1204648850288059e+01, + "cpu_time": 2.6558320615943046e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1505414782986115e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.8474656927025017e-03, + "cpu_time": 2.2589446615128399e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8557547814778498e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54128, + "real_time": 1.2809420665559028e+04, + "cpu_time": 1.1593581677906468e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2790586262852638e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54128, + "real_time": 1.2792955517966388e+04, + "cpu_time": 1.1529218598855164e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2807048361100264e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54128, + "real_time": 1.2915389122987026e+04, + "cpu_time": 1.1565373660902503e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2685641790567102e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54128, + "real_time": 1.2952926433759403e+04, + "cpu_time": 1.1606425621924317e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2648879065118549e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54128, + "real_time": 1.2828778803640213e+04, + "cpu_time": 1.1599892243507793e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2771285755859303e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2859894108782413e+04, + "cpu_time": 1.1578898360619249e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2740688247099571e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time_median", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2828778803640213e+04, + "cpu_time": 1.1593581677906468e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2771285755859303e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0233922924474101e+01, + "cpu_time": 3.1879026300592446e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9441922715086164e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.4614697703077681e-03, + "cpu_time": 2.7532002879492867e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4504059253545172e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53428, + "real_time": 1.3200250580917271e+04, + "cpu_time": 1.1603493521427534e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4823771184594431e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53428, + "real_time": 1.3249204477798898e+04, + "cpu_time": 1.1649085971892257e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4732050935516829e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53428, + "real_time": 1.3103653199931694e+04, + "cpu_time": 1.1650178371858335e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.5006766815357118e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53428, + "real_time": 1.3097723039559552e+04, + "cpu_time": 1.1607951079155152e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.5018088946475325e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53428, + "real_time": 1.3214989558461793e+04, + "cpu_time": 1.1614301479433744e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4796084669637942e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3173164171333841e+04, + "cpu_time": 1.1625002084753406e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4875352510316329e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time_median", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3200250580917269e+04, + "cpu_time": 1.1614301479433744e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4823771184594431e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8535217786765472e+01, + "cpu_time": 2.2813028593050987e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2954031693448955e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.2026390087740000e-03, + "cpu_time": 1.9624107098415978e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2075771340633860e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50608, + "real_time": 1.3693376002572959e+04, + "cpu_time": 1.1516060861995685e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7859636650367231e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50608, + "real_time": 1.3673856517302063e+04, + "cpu_time": 1.1444860846092544e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7927956474513788e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50608, + "real_time": 1.3820986549410109e+04, + "cpu_time": 1.1479553000870538e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7417743853312073e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50608, + "real_time": 1.3815324410387315e+04, + "cpu_time": 1.1528852656367731e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7437177769582825e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50608, + "real_time": 1.3689769907478527e+04, + "cpu_time": 1.1520051480931342e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7872243611778030e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3738662677430195e+04, + "cpu_time": 1.1497875769251568e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7702951671910791e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time_median", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3693376002572961e+04, + "cpu_time": 1.1516060861995685e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7859636650367231e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2964829646330543e+01, + "cpu_time": 3.5101044288417450e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5289146462540802e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3109120850748301e-03, + "cpu_time": 3.0528286261612904e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3013798048542893e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 46400, + "real_time": 1.5236637018500100e+04, + "cpu_time": 1.1469293460876973e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.6024232145751266e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 46400, + "real_time": 1.5236529348067837e+04, + "cpu_time": 1.1494047379873548e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.6024840044443207e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 46400, + "real_time": 1.5085611783821809e+04, + "cpu_time": 1.1500276788892889e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.6885438839520531e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 46400, + "real_time": 1.5086907600785424e+04, + "cpu_time": 1.1452665384187117e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.6877976234955120e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 46400, + "real_time": 1.5225671142639712e+04, + "cpu_time": 1.1471739573508017e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.6086188761118679e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5174271378762980e+04, + "cpu_time": 1.1477604517467709e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.6379735205157757e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time_median", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5225671142639714e+04, + "cpu_time": 1.1471739573508017e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.6086188761118679e+09, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0468032419565489e+01, + "cpu_time": 1.9428273300577472e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5893431031466804e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3029256173831073e-03, + "cpu_time": 1.6927115123203348e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3129858435507797e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 36141, + "real_time": 1.9170197357977137e+04, + "cpu_time": 1.0255814996938970e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3674559270560465e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 36141, + "real_time": 1.9263163181204898e+04, + "cpu_time": 1.0234075671482261e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3608564571356295e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 36141, + "real_time": 1.9351217901045813e+04, + "cpu_time": 1.0246514719323498e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3546640906039961e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 36141, + "real_time": 1.9373801811585727e+04, + "cpu_time": 1.0283657740041662e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3530849677797121e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 36141, + "real_time": 1.9174381029025230e+04, + "cpu_time": 1.0252037381612647e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3671575609308033e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9266552256167761e+04, + "cpu_time": 1.0254420101879807e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3606438007012375e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time_median", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9263163181204898e+04, + "cpu_time": 1.0252037381612647e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3608564571356295e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5475161569865023e+01, + "cpu_time": 1.8295173510223434e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7411469451431483e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.9554876399486971e-03, + "cpu_time": 1.7841256090990092e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9543803761638062e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 32222, + "real_time": 2.1878549845897454e+04, + "cpu_time": 9.4695277222797298e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.3963562653505192e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 32222, + "real_time": 2.1876404591541876e+04, + "cpu_time": 9.4779211492757571e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.3965912579743870e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 32222, + "real_time": 2.1773304125712966e+04, + "cpu_time": 9.4897238857992470e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.4079395436398067e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 32222, + "real_time": 2.1754405905676314e+04, + "cpu_time": 9.4603616868157160e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.4100313392755032e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 32222, + "real_time": 2.1877194463839485e+04, + "cpu_time": 9.4680270324320543e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.3965047294642300e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1831971786533621e+04, + "cpu_time": 9.4731122953204989e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.4014846271408894e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time_median", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1876404591541876e+04, + "cpu_time": 9.4695277222797275e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.3965912579743870e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2544465416793699e+01, + "cpu_time": 1.1184331612449721e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8876115715166166e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8648106560568353e-03, + "cpu_time": 1.1806396106984317e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8680639857839646e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 21748, + "real_time": 3.2051904613220941e+04, + "cpu_time": 7.5908912798190677e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2714935747296520e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 21748, + "real_time": 3.2044744713621119e+04, + "cpu_time": 7.5642861458121610e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2722245390654846e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 21748, + "real_time": 3.2192529038822315e+04, + "cpu_time": 7.5815031865288246e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2572029328154938e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 21748, + "real_time": 3.2198916722218335e+04, + "cpu_time": 7.5973947221912730e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2565567625958275e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 21748, + "real_time": 3.2058107481171159e+04, + "cpu_time": 7.5936264877363539e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2708605790777424e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2109240513810779e+04, + "cpu_time": 7.5855403644175363e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2656676776568405e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time_median", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2058107481171155e+04, + "cpu_time": 7.5908912798190677e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2708605790777424e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.9120954110655177e+01, + "cpu_time": 1.3253151492802550e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.0399069002848536e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4641178939322403e-03, + "cpu_time": 1.7471598404473331e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4619488857646385e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 13274, + "real_time": 5.2728827629091451e+04, + "cpu_time": 6.0659360495080996e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9772399544930588e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 13274, + "real_time": 5.2839201040443462e+04, + "cpu_time": 6.0701451822652268e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9689320782780693e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 13274, + "real_time": 5.2839830901358837e+04, + "cpu_time": 6.0795974326731896e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9688847678467293e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 13274, + "real_time": 5.2720195416040580e+04, + "cpu_time": 6.0792678555655968e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9778911732977440e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 13274, + "real_time": 5.2717291393017222e+04, + "cpu_time": 6.0676019611248147e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9781103023016518e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2769069275990303e+04, + "cpu_time": 6.0725096962273860e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9742116552434502e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time_median", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2728827629091451e+04, + "cpu_time": 6.0701451822652268e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9772399544930588e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4448943953209138e+01, + "cpu_time": 6.4961208980499278e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8517667049680777e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2213394103301556e-03, + "cpu_time": 1.0697588349814764e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2208123587395777e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 7458, + "real_time": 9.3939914891829278e+04, + "cpu_time": 5.0164355212079108e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4648794975274269e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 7458, + "real_time": 9.3962519391685302e+04, + "cpu_time": 5.0197468670261770e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4638053844809448e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 7458, + "real_time": 9.3822386867501787e+04, + "cpu_time": 5.0217967460807896e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4704724959974602e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 7458, + "real_time": 9.3826166360660151e+04, + "cpu_time": 5.0134432126348442e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4702924170187630e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 7458, + "real_time": 9.3936351303786461e+04, + "cpu_time": 5.0169116107279456e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4650488780810593e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.3897467763092602e+04, + "cpu_time": 5.0176667915355337e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4668997346211304e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time_median", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.3936351303786476e+04, + "cpu_time": 5.0169116107279462e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4650488780810593e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7576685547318732e+01, + "cpu_time": 3.2135720773584757e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2154684093826205e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.1968592079413307e-04, + "cpu_time": 6.4045147094652754e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1984342618233125e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3961, + "real_time": 1.7673336193549790e+05, + "cpu_time": 4.3851665780762208e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7464767874791962e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3961, + "real_time": 1.7663358982390969e+05, + "cpu_time": 4.3854768197929460e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7491578517782532e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3961, + "real_time": 1.7661169632968117e+05, + "cpu_time": 4.3818539735093129e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7497465764334091e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3961, + "real_time": 1.7673534310840300e+05, + "cpu_time": 4.3830518175513363e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7464235802879196e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3961, + "real_time": 1.7674116051664265e+05, + "cpu_time": 4.3851719679499263e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7462673524824432e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7669103034282688e+05, + "cpu_time": 4.3841442313759482e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7476144296922447e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time_median", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7673336193549790e+05, + "cpu_time": 4.3851665780762208e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7464767874791962e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2972053479353136e+01, + "cpu_time": 1.6059168864668987e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6922743983115740e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.5639643595473323e-04, + "cpu_time": 3.6630110728881919e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5644731124917246e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2050, + "real_time": 3.4123069285869418e+05, + "cpu_time": 4.0924589472538525e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9166784674166321e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2050, + "real_time": 3.4128563765368261e+05, + "cpu_time": 4.0923398505851251e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9158869137717918e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2050, + "real_time": 3.4142713464240020e+05, + "cpu_time": 4.0921035503363508e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9138496322420059e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2050, + "real_time": 3.4138290286802390e+05, + "cpu_time": 4.0936706216781777e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9144863023459457e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2050, + "real_time": 3.4124819679674128e+05, + "cpu_time": 4.0927834615317994e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9164262719879120e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4131491296390846e+05, + "cpu_time": 4.0926712862770617e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9154655175528587e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time_median", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4128563765368267e+05, + "cpu_time": 4.0924589472538531e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9158869137717918e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6048604024301795e+01, + "cpu_time": 6.1002379006541320e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2391346714245388e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5210912490483765e-04, + "cpu_time": 1.4905272067924303e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5208897651700672e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1042, + "real_time": 6.7107459284890722e+05, + "cpu_time": 3.9314713602686584e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0001046616221390e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1042, + "real_time": 6.7108825718109880e+05, + "cpu_time": 3.9314873804993063e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0000028522247040e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1042, + "real_time": 6.7122891359165276e+05, + "cpu_time": 3.9320747149613649e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9989550987091568e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1042, + "real_time": 6.7116580213082244e+05, + "cpu_time": 3.9329612313054597e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9994251634202942e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1042, + "real_time": 6.7107567086372513e+05, + "cpu_time": 3.9317798664875335e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0000966294625023e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7112664732324122e+05, + "cpu_time": 3.9319549107044649e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9997168810877594e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time_median", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7108825718109880e+05, + "cpu_time": 3.9317798664875329e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0000028522247040e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8508318791648989e+01, + "cpu_time": 6.1442659928522087e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1033656824648576e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0207956883382795e-04, + "cpu_time": 1.5626491484235704e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0207309341393243e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 526, + "real_time": 1.3299273728028280e+06, + "cpu_time": 3.8617180037263006e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0460547976065613e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 526, + "real_time": 1.3304201677129976e+06, + "cpu_time": 3.8616675746960479e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0441857112975555e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 526, + "real_time": 1.3301430432526078e+06, + "cpu_time": 3.8610437058554608e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0452366262727844e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 526, + "real_time": 1.3300752160564321e+06, + "cpu_time": 3.8612534288593894e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0454939081544937e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 526, + "real_time": 1.3303641929687797e+06, + "cpu_time": 3.8626723676048523e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0443979441631645e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3301859985587292e+06, + "cpu_time": 3.8616710161484104e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0450737974989128e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time_median", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3301430432526073e+06, + "cpu_time": 3.8616675746960479e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0452366262727844e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0469101172425289e+02, + "cpu_time": 6.2706095165084676e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.7634434388871538e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5388149623138254e-04, + "cpu_time": 1.6238072819477786e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5388166259799546e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 264, + "real_time": 2.6480359515654994e+06, + "cpu_time": 3.8350910662498903e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0685765017899948e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 264, + "real_time": 2.6486365840713861e+06, + "cpu_time": 3.8357813410604709e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0674270984238045e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 264, + "real_time": 2.6482402154766587e+06, + "cpu_time": 3.8351528055683517e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0681855526403618e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 264, + "real_time": 2.6481565895885453e+06, + "cpu_time": 3.8350777657199436e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0683456003957054e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 264, + "real_time": 2.6485978071860066e+06, + "cpu_time": 3.8357334174244052e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0675012882608688e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6483334295776198e+06, + "cpu_time": 3.8353672792046130e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0680072083021477e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time_median", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6482402154766587e+06, + "cpu_time": 3.8351528055683517e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0681855526403618e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6937547307364201e+02, + "cpu_time": 3.5763634310274785e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1548363670634590e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0171508997513372e-04, + "cpu_time": 9.3246960999499197e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0171328009595314e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 132, + "real_time": 5.2844572840540698e+06, + "cpu_time": 3.8612047323486608e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0797166401554245e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 132, + "real_time": 5.2840039380967170e+06, + "cpu_time": 3.8604967159850448e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0801524590969498e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 132, + "real_time": 5.2845615322108977e+06, + "cpu_time": 3.8610120624242854e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0796164329587227e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 132, + "real_time": 5.2845372684356393e+06, + "cpu_time": 3.8612329576514268e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0796397558468521e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 132, + "real_time": 5.2841723945951369e+06, + "cpu_time": 3.8607944784093183e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0799905066414284e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2843464834784921e+06, + "cpu_time": 3.8609481893637478e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0798231589398750e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time_median", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2844572840540698e+06, + "cpu_time": 3.8610120624242854e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0797166401554245e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4625488518076386e+02, + "cpu_time": 3.0758832207694726e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3672898740627435e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.6600821113959827e-05, + "cpu_time": 7.9666524125938931e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6601816637978814e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 66, + "real_time": 1.0556463377945350e+07, + "cpu_time": 3.9230969136363053e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0857080897153023e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 66, + "real_time": 1.0557021113169013e+07, + "cpu_time": 3.9237717812119484e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0854394080002151e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 66, + "real_time": 1.0555956696131915e+07, + "cpu_time": 3.9231030769699061e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0859522017244438e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 66, + "real_time": 1.0555353072105031e+07, + "cpu_time": 3.9227834507573873e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0862430496882751e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 66, + "real_time": 1.0557059847721547e+07, + "cpu_time": 3.9237756260603291e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0854207491858528e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0556370821414573e+07, + "cpu_time": 3.9233061697271758e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0857526996628181e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time_median", + "family_index": 1, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0556463377945350e+07, + "cpu_time": 3.9231030769699061e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0857080897153023e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2689694501740735e+02, + "cpu_time": 4.4594092071069063e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5020602572542923e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.8858602763634426e-05, + "cpu_time": 1.1366457304597798e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8860215273276582e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1098644049330190e+07, + "cpu_time": 4.0813164287877756e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0891508548583130e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1098842787923235e+07, + "cpu_time": 4.0811833630304980e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0891029180737770e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1104719376925267e+07, + "cpu_time": 4.0822365372728235e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0876858622151115e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1106920684828904e+07, + "cpu_time": 4.0827792772726452e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0871552512715759e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1106037960359544e+07, + "cpu_time": 4.0826936890911812e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0873680129669807e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1103032971873432e+07, + "cpu_time": 4.0820418590909851e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0880925798771515e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time_median", + "family_index": 1, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1104719376925267e+07, + "cpu_time": 4.0822365372728235e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0876858622151115e+10, + "dst_id": 1.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9940115121327594e+03, + "cpu_time": 7.5329810652789747e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6303304801611044e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8926244002253430e-04, + "cpu_time": 1.8453953500017432e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8927191926986560e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50468, + "real_time": 1.2763085580122250e+04, + "cpu_time": 1.1906908671409037e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0057845604256142e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50468, + "real_time": 1.2685050821482835e+04, + "cpu_time": 1.1722714073835654e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0181235660991587e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50468, + "real_time": 1.2692449706655369e+04, + "cpu_time": 1.1738956218258510e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0169471293296888e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50468, + "real_time": 1.2731429516543498e+04, + "cpu_time": 1.1756599847769706e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0107718435494460e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50468, + "real_time": 1.2643468032550099e+04, + "cpu_time": 1.1703052692521000e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0247609227226131e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2703096731470810e+04, + "cpu_time": 1.1765646300758777e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0152776044253048e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time_median", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2692449706655369e+04, + "cpu_time": 1.1738956218258510e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0169471293296888e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5814856793795357e+01, + "cpu_time": 8.1410060923546273e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.2670472446689106e+04, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6065896184426481e-03, + "cpu_time": 6.9193020801837334e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6059782675654005e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55019, + "real_time": 1.2755059843558007e+04, + "cpu_time": 1.2823345302685587e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0140932796845138e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55019, + "real_time": 1.2755251897070597e+04, + "cpu_time": 1.2791811329111969e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0140328402106054e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55019, + "real_time": 1.2713913438156174e+04, + "cpu_time": 1.2769921799614522e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0270842057443842e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55019, + "real_time": 1.2667154256359632e+04, + "cpu_time": 1.2777860938940289e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0419496726579040e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55019, + "real_time": 1.2700844397744235e+04, + "cpu_time": 1.2801276649801080e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0312280346567743e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2718444766577730e+04, + "cpu_time": 1.2792843204030688e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0256776065908365e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time_median", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2713913438156173e+04, + "cpu_time": 1.2791811329111969e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0270842057443842e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7604690784431043e+01, + "cpu_time": 2.0930986858435026e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1909692700957906e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9567051219384025e-03, + "cpu_time": 1.6361481591394965e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9584318131832925e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54552, + "real_time": 1.2755431694803621e+04, + "cpu_time": 1.2681714290985508e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0279525185898870e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54552, + "real_time": 1.2771522056181657e+04, + "cpu_time": 1.2691997340299020e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0178384024664059e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54552, + "real_time": 1.2767152018617009e+04, + "cpu_time": 1.2714725874263077e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0205828089679450e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54552, + "real_time": 1.2830786134303908e+04, + "cpu_time": 1.2719608201071589e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9808048336358145e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54552, + "real_time": 1.2822503104019954e+04, + "cpu_time": 1.2687709058365054e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9859602426531523e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2789479001585230e+04, + "cpu_time": 1.2699150952996852e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0066277612626418e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time_median", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2771522056181657e+04, + "cpu_time": 1.2691997340299020e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0178384024664059e+07, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4558051561143422e+01, + "cpu_time": 1.6935210956397362e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1616758686260067e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7020687517341421e-03, + "cpu_time": 1.3335703323064168e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6998580839295962e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54153, + "real_time": 1.2929967995909265e+04, + "cpu_time": 1.2665555928569720e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5839173002191025e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54153, + "real_time": 1.2983094632638102e+04, + "cpu_time": 1.2649226271549726e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5774359333802810e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54153, + "real_time": 1.2933926793081591e+04, + "cpu_time": 1.2641688244592221e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5834324971558392e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54153, + "real_time": 1.2898575698756837e+04, + "cpu_time": 1.2649534209280913e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5877722066610706e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54153, + "real_time": 1.2894057810873981e+04, + "cpu_time": 1.2659808027250278e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5883285386489075e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2927924586251955e+04, + "cpu_time": 1.2653162536248572e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5841772952130404e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time_median", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2929967995909265e+04, + "cpu_time": 1.2649534209280913e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5839173002191025e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5680007355636029e+01, + "cpu_time": 9.4614616892530699e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3653319256959145e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7599176586765909e-03, + "cpu_time": 7.4775469469771135e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7555829381514171e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53465, + "real_time": 1.3113098038582011e+04, + "cpu_time": 1.2531969908165298e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1235944305064631e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53465, + "real_time": 1.3097302120702285e+04, + "cpu_time": 1.2528695600686164e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1273616216927963e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53465, + "real_time": 1.3090876949171146e+04, + "cpu_time": 1.2541618535973208e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1288965711799312e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53465, + "real_time": 1.3108545981162013e+04, + "cpu_time": 1.2540670182829571e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1246791260344714e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53465, + "real_time": 1.3089022678041496e+04, + "cpu_time": 1.2518123985304489e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1293398298343253e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3099769153531790e+04, + "cpu_time": 1.2532215642591748e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1267743158495975e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time_median", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3097302120702285e+04, + "cpu_time": 1.2531969908165298e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1273616216927963e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0669303689073280e+01, + "cpu_time": 9.6296357623700588e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5461858243262608e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.1446501568287265e-04, + "cpu_time": 7.6839052542656237e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1431710994287702e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52986, + "real_time": 1.3229286437457979e+04, + "cpu_time": 1.2485511498136926e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1923218903211689e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52986, + "real_time": 1.3244475166162852e+04, + "cpu_time": 1.2452049303162429e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1852205521355975e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52986, + "real_time": 1.3230500502772580e+04, + "cpu_time": 1.2435327543724473e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1917536666759408e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52986, + "real_time": 1.3207182579335597e+04, + "cpu_time": 1.2463029594861741e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2026855090331519e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52986, + "real_time": 1.3195831492275915e+04, + "cpu_time": 1.2461567767291236e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2080210745303380e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3221455235600986e+04, + "cpu_time": 1.2459497141435361e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1960005385392392e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time_median", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3229286437457977e+04, + "cpu_time": 1.2461567767291234e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1923218903211689e+08, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9578985690968860e+01, + "cpu_time": 1.8256178552450556e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.1782160066104354e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4808495239048392e-03, + "cpu_time": 1.4652420033660688e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4813129775444274e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51418, + "real_time": 1.3607114785576350e+04, + "cpu_time": 1.2339071848288939e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2040759748251092e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51418, + "real_time": 1.3562756944172876e+04, + "cpu_time": 1.2320187553776538e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2080139803020835e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51418, + "real_time": 1.3545639254269148e+04, + "cpu_time": 1.2337202072805414e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2095405534173141e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51418, + "real_time": 1.3606723158869989e+04, + "cpu_time": 1.2328448188533909e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2041106303628697e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51418, + "real_time": 1.3575546650764760e+04, + "cpu_time": 1.2311126557702365e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2068758939498787e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3579556158730626e+04, + "cpu_time": 1.2327207244221435e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2065234065714509e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time_median", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3575546650764760e+04, + "cpu_time": 1.2328448188533909e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2068758939498787e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7139316416542886e+01, + "cpu_time": 1.1726784550337426e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4114644759398801e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9985422276923505e-03, + "cpu_time": 9.5129288556696705e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9986885151217098e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50116, + "real_time": 1.3989597139410193e+04, + "cpu_time": 1.2294063772694685e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3423119103043385e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50116, + "real_time": 1.4043332115531763e+04, + "cpu_time": 1.2259500229631188e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3333493597120709e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50116, + "real_time": 1.3998820507366549e+04, + "cpu_time": 1.2267499129783192e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3407686370974336e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50116, + "real_time": 1.3976625559914413e+04, + "cpu_time": 1.2313345176388409e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3444857887572012e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50116, + "real_time": 1.3979470462553718e+04, + "cpu_time": 1.2296338658827231e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3440086724153404e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3997569156955327e+04, + "cpu_time": 1.2286149393464942e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3409848736572771e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time_median", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3989597139410194e+04, + "cpu_time": 1.2294063772694685e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3423119103043385e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7044242987659192e+01, + "cpu_time": 2.2158967735932143e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5141063119071536e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9320671099682355e-03, + "cpu_time": 1.8035730338521357e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9282936693456071e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 46923, + "real_time": 1.4875266085450494e+04, + "cpu_time": 1.1993958102931542e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4057027029654818e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 46923, + "real_time": 1.4880438843786882e+04, + "cpu_time": 1.1991559527789841e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4041711866154823e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 46923, + "real_time": 1.4884570046980460e+04, + "cpu_time": 1.2037142861752079e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4029488116316051e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 46923, + "real_time": 1.4929181527583622e+04, + "cpu_time": 1.2037248096249597e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.3897918903935652e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 46923, + "real_time": 1.4862593292570748e+04, + "cpu_time": 1.2010077178123617e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4094592854639301e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4886409959274441e+04, + "cpu_time": 1.2013997153369336e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4024147754140120e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time_median", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4880438843786886e+04, + "cpu_time": 1.2010077178123617e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4041711866154823e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5297559613350209e+01, + "cpu_time": 2.2342263490203274e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.4689857193704685e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6993727623085830e-03, + "cpu_time": 1.8596860982223028e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6965656578026703e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 34724, + "real_time": 2.0279275727401713e+04, + "cpu_time": 1.0338584953460989e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4633472004571257e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 34724, + "real_time": 2.0318395995457900e+04, + "cpu_time": 1.0320686013015257e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4509029172037325e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 34724, + "real_time": 2.0247291390449285e+04, + "cpu_time": 1.0313467097582748e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4735572513085432e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 34724, + "real_time": 2.0171755655277724e+04, + "cpu_time": 1.0311892943632889e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4977983195878353e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 34724, + "real_time": 2.0261673002663782e+04, + "cpu_time": 1.0328611458497945e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4689623597601280e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0255678354250082e+04, + "cpu_time": 1.0322648493237965e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4709136096634731e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time_median", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0261673002663785e+04, + "cpu_time": 1.0320686013015258e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4689623597601280e+09, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3943569849889776e+01, + "cpu_time": 1.1105842172257493e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7257333267584536e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6631332165960881e-03, + "cpu_time": 1.0758713889688846e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6669083082507760e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 32331, + "real_time": 2.1628293604418875e+04, + "cpu_time": 1.0399679567052653e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2120419890473532e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 32331, + "real_time": 2.1625112430050591e+04, + "cpu_time": 1.0393436065333412e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2122202871681751e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 32331, + "real_time": 2.1673235249811649e+04, + "cpu_time": 1.0411652848294799e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2095286973931507e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 32331, + "real_time": 2.1650059455635845e+04, + "cpu_time": 1.0403217041650081e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2108234646522409e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 32331, + "real_time": 2.1639687543906115e+04, + "cpu_time": 1.0403819400584427e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2114038128698469e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1643277656764614e+04, + "cpu_time": 1.0402360984583075e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2112036502261536e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time_median", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1639687543906119e+04, + "cpu_time": 1.0403217041650079e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2114038128698469e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9434739347037542e+01, + "cpu_time": 6.6340546764247236e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0870183758108232e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.9795730828057859e-04, + "cpu_time": 6.3774509327803471e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9746953421735256e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 25789, + "real_time": 2.7108560832784548e+04, + "cpu_time": 9.0302305332323647e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9340311100762554e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 25789, + "real_time": 2.7147919154251660e+04, + "cpu_time": 9.0124271318180537e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9312272039011536e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 25789, + "real_time": 2.7143221846066539e+04, + "cpu_time": 9.0313486802033329e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9315614151235226e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 25789, + "real_time": 2.7139716120746096e+04, + "cpu_time": 9.0350300345482862e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9318109211880245e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 25789, + "real_time": 2.7130417499815325e+04, + "cpu_time": 9.0302980009459889e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9324730259074295e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7133967090732833e+04, + "cpu_time": 9.0278668761496067e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9322207352392773e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time_median", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7139716120746096e+04, + "cpu_time": 9.0302980009459901e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9318109211880245e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5583961769169779e+01, + "cpu_time": 8.8503777500020561e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1102772987653129e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.7433407054187174e-04, + "cpu_time": 9.8033986005969446e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7461204018588550e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 16938, + "real_time": 4.1317829249614922e+04, + "cpu_time": 7.1315341938470697e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5378293561000004e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 16938, + "real_time": 4.1254142810444879e+04, + "cpu_time": 7.1327382788154638e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5417471520812145e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 16938, + "real_time": 4.1309686645893606e+04, + "cpu_time": 7.1435379172983289e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5383295908012749e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 16938, + "real_time": 4.1317083398880495e+04, + "cpu_time": 7.1481544678417552e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5378751686727520e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 16938, + "real_time": 4.1334180600701395e+04, + "cpu_time": 7.1433291858001161e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5368254184823658e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1306584541107062e+04, + "cpu_time": 7.1398588087205470e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5385213372275215e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time_median", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1317083398880495e+04, + "cpu_time": 7.1433291858001161e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5378751686727520e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0651142544558230e+01, + "cpu_time": 7.3211760688329255e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8851498769318052e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.4204011019248374e-04, + "cpu_time": 1.0253950764251697e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.4261730610099806e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10053, + "real_time": 6.9632876178114559e+04, + "cpu_time": 5.6821813879993665e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0117268093819309e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10053, + "real_time": 6.9624409414102440e+04, + "cpu_time": 5.6798692071657515e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0120930542144337e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10053, + "real_time": 6.9679891043732670e+04, + "cpu_time": 5.6827056930214465e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0096947176392399e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10053, + "real_time": 6.9633449806275152e+04, + "cpu_time": 5.6769309877430344e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0117019993040924e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10053, + "real_time": 6.9626557015078593e+04, + "cpu_time": 5.6822160932511210e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0120001475095669e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9639436691460694e+04, + "cpu_time": 5.6807806738361442e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0114433456098530e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time_median", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9632876178114544e+04, + "cpu_time": 5.6821813879993665e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0117268093819309e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2951812944922704e+01, + "cpu_time": 2.4175908662320962e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9210670331935566e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.2958068065098371e-04, + "cpu_time": 4.2557370281284487e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2944558122458792e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5545, + "real_time": 1.2627856787707744e+05, + "cpu_time": 4.7947486771472198e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3214694072891571e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5545, + "real_time": 1.2629845132274475e+05, + "cpu_time": 4.7916565413219804e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3209465009842598e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5545, + "real_time": 1.2622461157722435e+05, + "cpu_time": 4.7909835144376618e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3228892112168797e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5545, + "real_time": 1.2624870447338642e+05, + "cpu_time": 4.7948905337353724e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3222550817415882e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5545, + "real_time": 1.2628200398184900e+05, + "cpu_time": 4.7943325721772754e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3213790308576855e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2626646784645638e+05, + "cpu_time": 4.7933223677639025e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3217878464179138e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time_median", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2627856787707742e+05, + "cpu_time": 4.7943325721772748e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3214694072891571e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9488174334840402e+01, + "cpu_time": 1.8546717042010112e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.7583971530619645e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3353923522038218e-04, + "cpu_time": 3.8692822261946478e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3356088684075104e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2916, + "real_time": 2.3996173763064481e+05, + "cpu_time": 4.2552504898525906e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4958106583275200e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2916, + "real_time": 2.3996223163223613e+05, + "cpu_time": 4.2506086217596602e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4958034616282036e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2916, + "real_time": 2.3992574151709329e+05, + "cpu_time": 4.2520504572330743e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4963351355954285e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2916, + "real_time": 2.3994741930459614e+05, + "cpu_time": 4.2492423174892956e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4960192630166443e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2916, + "real_time": 2.3987696227959779e+05, + "cpu_time": 4.2494716259876829e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4970461190943115e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3993481847283366e+05, + "cpu_time": 4.2513247024644613e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4962029275324219e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time_median", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3994741930459617e+05, + "cpu_time": 4.2506086217596602e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4960192630166443e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5583119455708193e+01, + "cpu_time": 2.4607680603182563e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1856334335238161e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4830327537367009e-04, + "cpu_time": 5.7882383316705207e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4832186635069762e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1499, + "real_time": 4.6704529936526663e+05, + "cpu_time": 3.9716539927156019e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.5922031594795868e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1499, + "real_time": 4.6707796541525994e+05, + "cpu_time": 3.9737016174720037e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.5919519314262794e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1499, + "real_time": 4.6687846890886163e+05, + "cpu_time": 3.9723335507537717e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.5934867673829369e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1499, + "real_time": 4.6698702035240055e+05, + "cpu_time": 3.9730918173048759e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.5926514589933304e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1499, + "real_time": 4.6692735829599958e+05, + "cpu_time": 3.9731814136027479e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.5931105132127228e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6698322246755764e+05, + "cpu_time": 3.9727924783697999e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.5926807660989716e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time_median", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6698702035240055e+05, + "cpu_time": 3.9730918173048753e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.5926514589933304e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.2064463936284511e+01, + "cpu_time": 8.0221138984341538e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.3136511154545117e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7573321692940629e-04, + "cpu_time": 2.0192632618268438e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7573649111913837e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 760, + "real_time": 9.2082056224554474e+05, + "cpu_time": 3.8401632669208270e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6439707556239845e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 760, + "real_time": 9.2078540252068231e+05, + "cpu_time": 3.8403156240521121e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6441098988041695e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 760, + "real_time": 9.2089235113235191e+05, + "cpu_time": 3.8407420543813419e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6436866870205452e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 760, + "real_time": 9.2073824412900175e+05, + "cpu_time": 3.8399670304205918e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6442965429052818e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 760, + "real_time": 9.2068834430967015e+05, + "cpu_time": 3.8397233093421304e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6444940578843788e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.2078498086745013e+05, + "cpu_time": 3.8401822570234007e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6441115884476723e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time_median", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.2078540252068231e+05, + "cpu_time": 3.8401632669208270e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6441098988041695e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7970399711980932e+01, + "cpu_time": 3.8356092238029414e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0857291613101754e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.4678183649918818e-05, + "cpu_time": 9.9880916245261654e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4677131487750134e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 383, + "real_time": 1.8286979346908140e+06, + "cpu_time": 3.7882130069973969e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6697621147227020e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 383, + "real_time": 1.8289623698146706e+06, + "cpu_time": 3.7886307472587270e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6692315330030632e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 383, + "real_time": 1.8283716778638079e+06, + "cpu_time": 3.7876364028197145e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6704169514596260e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 383, + "real_time": 1.8287113005679597e+06, + "cpu_time": 3.7879809168405557e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6697352927800789e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 383, + "real_time": 1.8286056115654330e+06, + "cpu_time": 3.7878430883289921e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6699473946462097e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8286697789005369e+06, + "cpu_time": 3.7880608324490774e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6698186573223358e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time_median", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8286979346908145e+06, + "cpu_time": 3.7879809168405557e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6697621147227020e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1278387813242225e+02, + "cpu_time": 3.8141168960688941e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2702191248318860e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1635992489598403e-04, + "cpu_time": 1.0068784702179588e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1636049417078362e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 192, + "real_time": 3.6438255859441902e+06, + "cpu_time": 3.7919540259374374e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6834289906118385e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 192, + "real_time": 3.6439347544122334e+06, + "cpu_time": 3.7916649651044011e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6833186389378510e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 192, + "real_time": 3.6440464561261856e+06, + "cpu_time": 3.7924503246869534e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6832057334055107e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 192, + "real_time": 3.6438981296669226e+06, + "cpu_time": 3.7925979673954064e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6833556598978912e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 192, + "real_time": 3.6437397454089173e+06, + "cpu_time": 3.7912965910414147e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6835157661606667e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6438889343116903e+06, + "cpu_time": 3.7919927748331225e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6833649578027519e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time_median", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6438981296669231e+06, + "cpu_time": 3.7919540259374368e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6833556598978912e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1539875726655582e+02, + "cpu_time": 5.4065758124117470e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1664866908799261e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1669120367510316e-05, + "cpu_time": 1.4257874773112349e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1669050019300115e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 96, + "real_time": 7.2738529658333091e+06, + "cpu_time": 3.8187162020837802e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6904163070231575e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 96, + "real_time": 7.2741179901640862e+06, + "cpu_time": 3.8187985720836669e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6902818508439507e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 96, + "real_time": 7.2742146294331178e+06, + "cpu_time": 3.8186737747916412e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6902328247760170e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 96, + "real_time": 7.2742862976156175e+06, + "cpu_time": 3.8190259481253600e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6901964676313110e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 96, + "real_time": 7.2744712670100853e+06, + "cpu_time": 3.8195050252081347e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6901026362886566e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2741886300112437e+06, + "cpu_time": 3.8189439044585168e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6902460173126183e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time_median", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2742146294331178e+06, + "cpu_time": 3.8187985720836675e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6902328247760170e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2790297647802964e+02, + "cpu_time": 3.4187993952999575e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1561787322036331e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1330363848108977e-05, + "cpu_time": 8.9522116083155838e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1330668112084509e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 48, + "real_time": 1.4533762199183306e+07, + "cpu_time": 3.9155761435413909e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6939569028463142e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 48, + "real_time": 1.4533388758233437e+07, + "cpu_time": 3.9153156887498379e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6940518204734085e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 48, + "real_time": 1.4535302199268093e+07, + "cpu_time": 3.9159156170837402e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6935655319710754e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 48, + "real_time": 1.4535078972888490e+07, + "cpu_time": 3.9160278358332807e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6936222568958641e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 48, + "real_time": 1.4534295575382808e+07, + "cpu_time": 3.9154613199999720e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6938213428748146e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4534365540991228e+07, + "cpu_time": 3.9156593210416448e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6938035710122955e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time_median", + "family_index": 2, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4534295575382810e+07, + "cpu_time": 3.9155761435413915e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6938213428748146e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.2300286364450164e+02, + "cpu_time": 3.0236437780000939e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0916002922164646e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.6624615730448580e-05, + "cpu_time": 7.7219275991450228e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6624567387140640e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 24, + "real_time": 2.9056457647432882e+07, + "cpu_time": 4.1370562020829028e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6953638224887482e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 24, + "real_time": 2.9057284351438284e+07, + "cpu_time": 4.1369276529166633e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6952586863020172e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 24, + "real_time": 2.9057751021658380e+07, + "cpu_time": 4.1371470245830721e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6951993400992378e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 24, + "real_time": 2.9060057519624632e+07, + "cpu_time": 4.1374974829169750e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6949060519749084e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 24, + "real_time": 2.9056890790040296e+07, + "cpu_time": 4.1368538912497377e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6953087367766197e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9057688266038902e+07, + "cpu_time": 4.1370964507498705e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6952073275283066e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time_median", + "family_index": 2, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9057284351438284e+07, + "cpu_time": 4.1370562020829028e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6952586863020172e+10, + "dst_id": 2.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4080871082483736e+03, + "cpu_time": 2.5116342597645602e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7905515517627522e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.8458332106689713e-05, + "cpu_time": 6.0710072623743472e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8456051123941599e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57814, + "real_time": 1.2242721881805588e+04, + "cpu_time": 1.1950887074194059e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0910382713214461e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57814, + "real_time": 1.2234395124336808e+04, + "cpu_time": 1.2008169926127276e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0924614367796712e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57814, + "real_time": 1.2023273240411503e+04, + "cpu_time": 1.1950515721245725e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1292038771900877e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57814, + "real_time": 1.2052559666001347e+04, + "cpu_time": 1.1964348952431467e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1240301404368203e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57814, + "real_time": 1.2250903271756973e+04, + "cpu_time": 1.1946042046423278e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0896418355549186e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2160770636862444e+04, + "cpu_time": 1.1963992744084363e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1052751122565892e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time_median", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2234395124336808e+04, + "cpu_time": 1.1950887074194059e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0924614367796712e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1277811381443614e+02, + "cpu_time": 2.5628922002602033e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9593456386954803e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.2739282058800169e-03, + "cpu_time": 2.1421713094296507e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.3068389365763664e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56988, + "real_time": 1.2073412147681085e+04, + "cpu_time": 1.1797341357195740e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2407232830061115e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56988, + "real_time": 1.2106047606838227e+04, + "cpu_time": 1.1818202608051021e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2292911495804086e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56988, + "real_time": 1.2301418032141757e+04, + "cpu_time": 1.1809532607968841e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1621217867909282e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56988, + "real_time": 1.2286010115958365e+04, + "cpu_time": 1.1837613643837030e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1673415141906850e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56988, + "real_time": 1.2075245624380219e+04, + "cpu_time": 1.1815108066569555e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2400793816256568e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2168426705399932e+04, + "cpu_time": 1.1815559656724439e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2079114230387583e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time_median", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2106047606838227e+04, + "cpu_time": 1.1815108066569552e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2292911495804086e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1523257287214058e+02, + "cpu_time": 1.4678208567817935e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9721205648911576e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.4698004649199485e-03, + "cpu_time": 1.2422778940872525e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.4396487130013699e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57518, + "real_time": 1.2390521673885123e+04, + "cpu_time": 1.1961037084749663e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2643816535847157e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57518, + "real_time": 1.2341154216750321e+04, + "cpu_time": 1.1987357425103812e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2974410822137848e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57518, + "real_time": 1.2147753819552991e+04, + "cpu_time": 1.1945952140975864e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.4295419154096827e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57518, + "real_time": 1.2201241015413983e+04, + "cpu_time": 1.1971447741404462e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.3925889071969628e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57518, + "real_time": 1.2363458178080868e+04, + "cpu_time": 1.1938331676871898e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2824723087222144e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2288825780736657e+04, + "cpu_time": 1.1960825213821142e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.3332851734254733e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time_median", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2341154216750323e+04, + "cpu_time": 1.1961037084749665e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2974410822137848e+07, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0749733017731828e+02, + "cpu_time": 1.9646938166721244e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3138113109975017e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.7475672692687752e-03, + "cpu_time": 1.6426072461972387e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7766242949670857e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56238, + "real_time": 1.2343351218495091e+04, + "cpu_time": 1.1770249567201855e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6591928429706416e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56238, + "real_time": 1.2279437332416588e+04, + "cpu_time": 1.1736565257160511e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6678288626413426e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56238, + "real_time": 1.2489054968058153e+04, + "cpu_time": 1.1724409303399866e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6398358444557562e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56238, + "real_time": 1.2448641081370162e+04, + "cpu_time": 1.1739124111756985e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6451594889862359e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56238, + "real_time": 1.2275377117888314e+04, + "cpu_time": 1.1727412296281569e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6683805151823390e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2367172343645663e+04, + "cpu_time": 1.1739552107160163e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6560795108472633e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time_median", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2343351218495089e+04, + "cpu_time": 1.1736565257160511e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6591928429706416e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7703340565653463e+01, + "cpu_time": 1.8221676785913887e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3059081011139336e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.9002166259819380e-03, + "cpu_time": 1.5521611573920411e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8855398702796630e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56317, + "real_time": 1.2664253878811567e+04, + "cpu_time": 1.1795181562969048e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2343002905626959e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56317, + "real_time": 1.2617480105449235e+04, + "cpu_time": 1.1824196298438547e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2462900402997428e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56317, + "real_time": 1.2428885966091810e+04, + "cpu_time": 1.1806883799038143e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2955487814230573e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56317, + "real_time": 1.2459996715719075e+04, + "cpu_time": 1.1816969140113947e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2873202886423206e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56317, + "real_time": 1.2640643379138111e+04, + "cpu_time": 1.1793380415021036e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2403413949324477e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2562252009041960e+04, + "cpu_time": 1.1807322243116145e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2607601591720527e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time_median", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2617480105449236e+04, + "cpu_time": 1.1806883799038143e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2462900402997428e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0936450223846080e+02, + "cpu_time": 1.3414164928930751e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8469824532687236e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.7058038765456414e-03, + "cpu_time": 1.1360886619954343e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7310391267526027e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54808, + "real_time": 1.2602007289690264e+04, + "cpu_time": 1.1561788530058053e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.5005517071093082e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54808, + "real_time": 1.2592554418855156e+04, + "cpu_time": 1.1550618892935607e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.5054314855561852e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54808, + "real_time": 1.2839541286375914e+04, + "cpu_time": 1.1532310141155429e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.3802902434626400e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54808, + "real_time": 1.2802862933503644e+04, + "cpu_time": 1.1549597951055090e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.3985688533479989e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54808, + "real_time": 1.2601423025683978e+04, + "cpu_time": 1.1533954792310822e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.5008531046876395e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2687677790821794e+04, + "cpu_time": 1.1545654061502998e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4571390788327551e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time_median", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2602007289690264e+04, + "cpu_time": 1.1549597951055090e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.5005517071093082e+08, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2263555106884981e+02, + "cpu_time": 1.2404213008585332e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2177053059211485e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.6657207954605993e-03, + "cpu_time": 1.0743620883242161e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6291952674575371e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53969, + "real_time": 1.3166593384729484e+04, + "cpu_time": 1.1580882035861788e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2443613561425872e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53969, + "real_time": 1.3139504226881882e+04, + "cpu_time": 1.1591178806223168e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2469268031042039e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53969, + "real_time": 1.2922673842707076e+04, + "cpu_time": 1.1584421058705933e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2678490689638760e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53969, + "real_time": 1.2982426504345252e+04, + "cpu_time": 1.1591590279097307e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2620136917021043e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53969, + "real_time": 1.3172304694147131e+04, + "cpu_time": 1.1590192611149731e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2438218201314404e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3076700530562168e+04, + "cpu_time": 1.1587652958207586e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2529945480088422e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time_median", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3139504226881883e+04, + "cpu_time": 1.1590192611149731e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2469268031042039e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1594934043288939e+02, + "cpu_time": 4.7611579664217104e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1152244308959519e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.8668651669355572e-03, + "cpu_time": 4.1088199513684612e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9004731318917275e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51793, + "real_time": 1.3317563646017856e+04, + "cpu_time": 1.1411520344899626e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4605101106310921e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51793, + "real_time": 1.3341041283580842e+04, + "cpu_time": 1.1419335491558273e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4561800914542112e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51793, + "real_time": 1.3551704615614832e+04, + "cpu_time": 1.1404075172447810e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4179983942568641e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51793, + "real_time": 1.3525230193588990e+04, + "cpu_time": 1.1423822811249838e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4227314087069774e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51793, + "real_time": 1.3329364464608720e+04, + "cpu_time": 1.1402800526811066e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4583317597026858e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3412980840682247e+04, + "cpu_time": 1.1412310869393322e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4431503529503665e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time_median", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3341041283580842e+04, + "cpu_time": 1.1411520344899626e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4561800914542112e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1523415533258165e+02, + "cpu_time": 9.2299562723755813e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0923459879520882e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.5912413281819238e-03, + "cpu_time": 8.0877189361616524e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.5641310835632995e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49350, + "real_time": 1.4405327844589621e+04, + "cpu_time": 1.1440514206545141e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5494278718976974e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49350, + "real_time": 1.4364711516340534e+04, + "cpu_time": 1.1455516845090368e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5622914129148865e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49350, + "real_time": 1.4154530266918662e+04, + "cpu_time": 1.1425102414797511e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6300370810021029e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49350, + "real_time": 1.4183993908997036e+04, + "cpu_time": 1.1437978941360176e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6204193558226166e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49350, + "real_time": 1.4410581535426723e+04, + "cpu_time": 1.1437304322173378e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5477692790459175e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4303829014454517e+04, + "cpu_time": 1.1439283345993314e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5819890001366444e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time_median", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4364711516340534e+04, + "cpu_time": 1.1437978941360173e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5622914129148865e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2455456716244532e+02, + "cpu_time": 1.0859999158304674e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0014752401695050e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.7077779688626445e-03, + "cpu_time": 9.4936009799149372e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7330529166485833e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 42510, + "real_time": 1.6258626129128834e+04, + "cpu_time": 1.0903909803278692e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.0616897737240162e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 42510, + "real_time": 1.6344994641950720e+04, + "cpu_time": 1.0921923234980290e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.0190910349761372e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 42510, + "real_time": 1.6542614391870156e+04, + "cpu_time": 1.0915412483885906e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.9232941598647871e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 42510, + "real_time": 1.6502379628542138e+04, + "cpu_time": 1.0917794035727112e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.9426120929433031e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 42510, + "real_time": 1.6227627936291159e+04, + "cpu_time": 1.0893508892543509e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.0770893019350710e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6375248545556602e+04, + "cpu_time": 1.0910509690083101e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.0047552726886635e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time_median", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6344994641950718e+04, + "cpu_time": 1.0915412483885906e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.0190910349761372e+09, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4184636913747318e+02, + "cpu_time": 1.1618452011630861e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9240434297535717e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.6622422092007233e-03, + "cpu_time": 1.0648862740290886e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.6499127004889698e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 36381, + "real_time": 1.9381295131062481e+04, + "cpu_time": 1.0462070341468616e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3525618294716576e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 36381, + "real_time": 1.9357801502211685e+04, + "cpu_time": 1.0461946002056495e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3542033684457882e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 36381, + "real_time": 1.9127249970583176e+04, + "cpu_time": 1.0449122441829382e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3705263454138222e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 36381, + "real_time": 1.9241498625688830e+04, + "cpu_time": 1.0458197345264634e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3623886844760536e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 36381, + "real_time": 1.9386734852847654e+04, + "cpu_time": 1.0457594202208519e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3521823142977299e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9298916016478765e+04, + "cpu_time": 1.0457786066565529e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3583725084210106e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time_median", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9357801502211682e+04, + "cpu_time": 1.0458197345264635e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3542033684457882e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1261877624249628e+02, + "cpu_time": 5.2660333802012028e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.9578521759885937e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.8354975039185868e-03, + "cpu_time": 5.0355145407278683e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8583725205384948e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27726, + "real_time": 2.5060369780206678e+04, + "cpu_time": 8.7265246675453079e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0921000152762951e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27726, + "real_time": 2.5111874728575800e+04, + "cpu_time": 8.7342793055689323e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0878090770474888e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27726, + "real_time": 2.5304693621282207e+04, + "cpu_time": 8.7292761701680636e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0719002088966366e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27726, + "real_time": 2.5241927576006932e+04, + "cpu_time": 8.7304038590254891e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0770521523021423e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27726, + "real_time": 2.5064433485525675e+04, + "cpu_time": 8.7240039273862064e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0917608223731377e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5156659838319461e+04, + "cpu_time": 8.7288975859388006e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0841244551791401e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time_median", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5111874728575800e+04, + "cpu_time": 8.7292761701680636e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0878090770474888e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1064331146138775e+02, + "cpu_time": 3.9036433790160721e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.1508164424963966e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.3981717832369850e-03, + "cpu_time": 4.4720920833168785e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3907245652994569e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 17910, + "real_time": 3.9250136168539997e+04, + "cpu_time": 6.9973300420190227e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6715219419810852e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 17910, + "real_time": 3.9218690222561672e+04, + "cpu_time": 6.9964233606092000e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6736639955323563e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 17910, + "real_time": 3.8985256039239335e+04, + "cpu_time": 6.9968907168266082e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6896732419676559e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 17910, + "real_time": 3.9073448223325438e+04, + "cpu_time": 6.9948966839426184e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6836024146290680e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 17910, + "real_time": 3.9267270469813542e+04, + "cpu_time": 6.9983408236847782e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6703562214900723e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9158960224695998e+04, + "cpu_time": 6.9967763254164445e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6777635631200478e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time_median", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9218690222561672e+04, + "cpu_time": 6.9968907168266082e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6736639955323563e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2358932093969329e+02, + "cpu_time": 1.2673770711712111e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4641248229374304e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1560930175503080e-03, + "cpu_time": 1.8113728554782369e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1608932691112178e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10434, + "real_time": 6.6917308981388487e+04, + "cpu_time": 5.5504573669584095e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1339455096487434e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10434, + "real_time": 6.6740410067391203e+04, + "cpu_time": 5.5514993172692978e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1422521945585869e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10434, + "real_time": 6.6879356725268415e+04, + "cpu_time": 5.5522838720515978e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1357239403704552e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10434, + "real_time": 6.6957454826458852e+04, + "cpu_time": 5.5490716496833324e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1320664822691124e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10434, + "real_time": 6.6948300211567912e+04, + "cpu_time": 5.5526874789941931e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1324947659203392e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6888566162414994e+04, + "cpu_time": 5.5511999369913650e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1352965785534481e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time_median", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6917308981388487e+04, + "cpu_time": 5.5514993172692978e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1339455096487434e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8282212996308800e+01, + "cpu_time": 1.4627322149206937e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1434374418465644e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3198401170978456e-03, + "cpu_time": 2.6349838440758153e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3215456139585521e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5692, + "real_time": 1.2276851501990476e+05, + "cpu_time": 4.6777836420638460e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4164329505166428e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5692, + "real_time": 1.2282656144140476e+05, + "cpu_time": 4.6770468426916349e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4148183835634945e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5692, + "real_time": 1.2303159378407388e+05, + "cpu_time": 4.6777868500934827e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4091275834085323e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5692, + "real_time": 1.2297093321275403e+05, + "cpu_time": 4.6786051039179689e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4108092785986797e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5692, + "real_time": 1.2278105293549315e+05, + "cpu_time": 4.6784077435457808e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4160840778940117e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2287573127872610e+05, + "cpu_time": 4.6779260364625424e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4134544547962727e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time_median", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2282656144140474e+05, + "cpu_time": 4.6777868500934827e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4148183835634945e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1856803590489778e+02, + "cpu_time": 6.1355623261434550e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2926216450200286e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.6494266744946624e-04, + "cpu_time": 1.3115988321147506e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6460101888675830e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2985, + "real_time": 2.3453914455272310e+05, + "cpu_time": 4.2015664353872782e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5766345170216515e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2985, + "real_time": 2.3467529900886829e+05, + "cpu_time": 4.2045818946095479e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5745594169597702e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2985, + "real_time": 2.3464812342174212e+05, + "cpu_time": 4.2025631941937429e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5749734017360245e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2985, + "real_time": 2.3445709001751058e+05, + "cpu_time": 4.2054797904923218e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5778862560195946e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2985, + "real_time": 2.3449332386717806e+05, + "cpu_time": 4.2007497725759017e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5773334019315125e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3456259617360443e+05, + "cpu_time": 4.2029882174517596e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5762773987337105e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time_median", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3453914455272304e+05, + "cpu_time": 4.2025631941937423e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5766345170216515e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5520965162069103e+01, + "cpu_time": 1.9975289466342723e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4562767215190938e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.0723016678828079e-04, + "cpu_time": 4.7526398916372881e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0720463184280186e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1528, + "real_time": 4.5820220763044170e+05, + "cpu_time": 3.9531497378535986e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6615310272646027e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1528, + "real_time": 4.5810461277294543e+05, + "cpu_time": 3.9534109574020660e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6623110818391701e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1528, + "real_time": 4.5789665708556672e+05, + "cpu_time": 3.9533091776375216e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6639743357778778e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1528, + "real_time": 4.5791487423587264e+05, + "cpu_time": 3.9520527873496532e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6638285725040741e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1528, + "real_time": 4.5813634326219489e+05, + "cpu_time": 3.9533679878531122e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6620574304445152e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5805093899740436e+05, + "cpu_time": 3.9530581296191895e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6627404895660484e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time_median", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5810461277294549e+05, + "cpu_time": 3.9533091776375222e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6623110818391701e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3727116360336805e+02, + "cpu_time": 5.7066405692358443e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0977318387992579e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9968536666212563e-04, + "cpu_time": 1.4436014807061749e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9970232450984103e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 773, + "real_time": 9.0475696022170479e+05, + "cpu_time": 3.8279029247735929e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7086680153063103e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 773, + "real_time": 9.0470686609125463e+05, + "cpu_time": 3.8283704700134563e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7088733663501877e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 773, + "real_time": 9.0452202777697542e+05, + "cpu_time": 3.8277704560419822e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7096312714977234e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 773, + "real_time": 9.0457169946890324e+05, + "cpu_time": 3.8277032649672240e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7094275688373459e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 773, + "real_time": 9.0477185490701394e+05, + "cpu_time": 3.8280090520573688e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7086069618565315e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0466588169317064e+05, + "cpu_time": 3.8279512335707253e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7090414367696190e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time_median", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0470686609125475e+05, + "cpu_time": 3.8279029247735929e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7088733663501877e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1266053124437376e+02, + "cpu_time": 2.6249038907790753e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6191523001347333e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2453275128881678e-04, + "cpu_time": 6.8572030588032236e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2453762997476171e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 389, + "real_time": 1.7978575180588737e+06, + "cpu_time": 3.7779243823646307e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7327131502866081e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 389, + "real_time": 1.7982848855884683e+06, + "cpu_time": 3.7788616829302293e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7318260603652565e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 389, + "real_time": 1.7976777772909955e+06, + "cpu_time": 3.7780907809516841e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7330863655181564e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 389, + "real_time": 1.7976855762197534e+06, + "cpu_time": 3.7780086981750911e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7330701702084778e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 389, + "real_time": 1.7982642827699184e+06, + "cpu_time": 3.7788800095886189e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7318688161136292e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7979540079856019e+06, + "cpu_time": 3.7783531108020508e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7325129124984253e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time_median", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7978575180588732e+06, + "cpu_time": 3.7780907809516841e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7327131502866081e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0142152524953639e+02, + "cpu_time": 4.7631738992060862e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2572072541542044e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6764696088485828e-04, + "cpu_time": 1.2606481605936988e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6764060569493996e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 195, + "real_time": 3.5841912747575687e+06, + "cpu_time": 3.7855663648208749e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7447144337763718e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 195, + "real_time": 3.5847426810039161e+06, + "cpu_time": 3.7861010258971667e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7441384206247124e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 195, + "real_time": 3.5839065467604459e+06, + "cpu_time": 3.7853848044102657e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7450119373598534e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 195, + "real_time": 3.5839672654103008e+06, + "cpu_time": 3.7849086464617789e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7449484903326660e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 195, + "real_time": 3.5847262455484807e+06, + "cpu_time": 3.7863518587181693e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7441555869620956e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5843068026961423e+06, + "cpu_time": 3.7856625400616515e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7445937738111404e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time_median", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5841912747575687e+06, + "cpu_time": 3.7855663648208749e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7447144337763718e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0458598948974685e+02, + "cpu_time": 5.7492876856829869e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2266937737763785e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1287705315443819e-04, + "cpu_time": 1.5187005246350765e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1287456074239451e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 98, + "real_time": 7.1567920445255479e+06, + "cpu_time": 3.8347474680612975e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7507790408041069e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 98, + "real_time": 7.1568377743646195e+06, + "cpu_time": 3.8345267310205483e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7507550745598892e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 98, + "real_time": 7.1571512641955391e+06, + "cpu_time": 3.8350708991831952e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7505907880259399e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 98, + "real_time": 7.1564655684466874e+06, + "cpu_time": 3.8348389131633157e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7509501503584259e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 98, + "real_time": 7.1563039602217628e+06, + "cpu_time": 3.8344990945923924e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7510348567094910e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1567101223508315e+06, + "cpu_time": 3.8347366212041497e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7508219820915703e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time_median", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1567920445255470e+06, + "cpu_time": 3.8347474680612981e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7507790408041069e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3256078763596560e+02, + "cpu_time": 2.3600628305120295e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7429428693333582e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.6468388680066617e-05, + "cpu_time": 6.1544326602825289e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6468290888106648e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49, + "real_time": 1.4301172742734149e+07, + "cpu_time": 3.9307871424490970e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7540341736852486e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49, + "real_time": 1.4302893598773042e+07, + "cpu_time": 3.9317487999998397e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7535825061724213e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49, + "real_time": 1.4301097742757019e+07, + "cpu_time": 3.9308584755105948e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7540538611583534e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49, + "real_time": 1.4302129686182858e+07, + "cpu_time": 3.9311145185710144e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7537829944212120e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49, + "real_time": 1.4301440849596141e+07, + "cpu_time": 3.9312826022448307e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7539637973971046e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4301746924008643e+07, + "cpu_time": 3.9311583077550757e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7538834665668678e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time_median", + "family_index": 3, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4301440849596143e+07, + "cpu_time": 3.9311145185710144e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7539637973971046e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5930425731071205e+02, + "cpu_time": 3.8526313598978595e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9929364057691353e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3091713994466798e-05, + "cpu_time": 9.8002447581357761e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3089991298845113e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 24, + "real_time": 2.8594980482012033e+07, + "cpu_time": 4.0740442604173189e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7550010732668564e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 24, + "real_time": 2.8593106971432764e+07, + "cpu_time": 4.0736182358335781e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7552471127841064e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 24, + "real_time": 2.8594553858662646e+07, + "cpu_time": 4.0741889354166478e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7550570969118752e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 24, + "real_time": 2.8592453648646671e+07, + "cpu_time": 4.0735669887502956e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7553329182395020e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 24, + "real_time": 2.8591147080684703e+07, + "cpu_time": 4.0732607666663033e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7555045307202347e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8593248408287771e+07, + "cpu_time": 4.0737358374168289e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7552285463845154e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time_median", + "family_index": 3, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8593106971432768e+07, + "cpu_time": 4.0736182358335781e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7552471127841064e+10, + "dst_id": 3.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5629702581348788e+03, + "cpu_time": 3.7697970007590637e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0527084275756262e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.4662213814148203e-05, + "cpu_time": 9.2539063680415422e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4662676378297852e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55235, + "real_time": 1.2694458552128202e+04, + "cpu_time": 1.2757904153330507e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0166279558026686e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55235, + "real_time": 1.2707951422811680e+04, + "cpu_time": 1.2772822231173289e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0144867688151661e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55235, + "real_time": 1.2695909818193857e+04, + "cpu_time": 1.2763980594049513e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0163974355988219e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55235, + "real_time": 1.2700312603354543e+04, + "cpu_time": 1.2763750394485309e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0156984162136491e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55235, + "real_time": 1.2704903649802016e+04, + "cpu_time": 1.2751947816139512e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0149700230429478e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2700707209258060e+04, + "cpu_time": 1.2762081037835627e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0156361198946506e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time_median", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2700312603354543e+04, + "cpu_time": 1.2763750394485307e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0156984162136491e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.7511984393228417e+00, + "cpu_time": 7.7765166675625718e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.1267653429350321e+03, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.5282505490171129e-04, + "cpu_time": 6.0934550129462448e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5279826318115653e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54869, + "real_time": 1.2728886824142292e+04, + "cpu_time": 1.2699541854200830e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0223470211779498e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54869, + "real_time": 1.2737529606111721e+04, + "cpu_time": 1.2708843623065078e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0196177424728587e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54869, + "real_time": 1.2732074112131510e+04, + "cpu_time": 1.2699417898818119e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0213400856043614e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54869, + "real_time": 1.2759744141040146e+04, + "cpu_time": 1.2715191421394420e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0126196445680678e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54869, + "real_time": 1.2713790435631679e+04, + "cpu_time": 1.2694573323245804e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0271231667077690e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2734405023811471e+04, + "cpu_time": 1.2703513624144847e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0206095321062014e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time_median", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2732074112131511e+04, + "cpu_time": 1.2699541854200830e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0213400856043614e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6678420182437360e+01, + "cpu_time": 8.3255441274099727e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2631255438783708e+04, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3097133435956499e-03, + "cpu_time": 6.5537333793904719e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3090367273544407e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54590, + "real_time": 1.2821073300662985e+04, + "cpu_time": 1.2697492298393130e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9868508352342740e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54590, + "real_time": 1.2832566273358734e+04, + "cpu_time": 1.2697165511094325e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9796977329927579e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54590, + "real_time": 1.2782807282207645e+04, + "cpu_time": 1.2654460374678738e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0107599011157975e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54590, + "real_time": 1.2828090023564524e+04, + "cpu_time": 1.2701469897158074e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9824821787106737e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54590, + "real_time": 1.2787985885595006e+04, + "cpu_time": 1.2663647062370052e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0075158759244651e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2810504553077779e+04, + "cpu_time": 1.2682847028738866e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9934613047955945e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time_median", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2821073300662983e+04, + "cpu_time": 1.2697165511094327e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9868508352342740e+07, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3355384273500619e+01, + "cpu_time": 2.2027000053864894e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4581219345445701e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8231432007015981e-03, + "cpu_time": 1.7367551626186555e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8241433578589851e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54025, + "real_time": 1.2936129396229848e+04, + "cpu_time": 1.2596332943540239e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5831628899730060e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54025, + "real_time": 1.2957837975437629e+04, + "cpu_time": 1.2598555540613573e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5805105789114732e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54025, + "real_time": 1.2968828709035977e+04, + "cpu_time": 1.2592125276493173e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5791711386958674e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54025, + "real_time": 1.2962973311221369e+04, + "cpu_time": 1.2600145842611167e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5798844530730873e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54025, + "real_time": 1.2947949830220638e+04, + "cpu_time": 1.2574047640637193e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5817175899306843e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2954743844429093e+04, + "cpu_time": 1.2592241448779070e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5808893301168236e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time_median", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2957837975437631e+04, + "cpu_time": 1.2596332943540237e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5805105789114732e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2926112062667988e+01, + "cpu_time": 1.0607964902375389e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5780130582476180e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.9778986121957050e-04, + "cpu_time": 8.4242070369480769e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9818059884748981e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53664, + "real_time": 1.3074344684304737e+04, + "cpu_time": 1.2560252365292780e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1328530025042826e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53664, + "real_time": 1.3083003508800815e+04, + "cpu_time": 1.2558916116359322e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1307795623876876e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53664, + "real_time": 1.3061249482575815e+04, + "cpu_time": 1.2540687904103200e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1359939992450291e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53664, + "real_time": 1.3101403341869105e+04, + "cpu_time": 1.2552527914537945e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1263826424686247e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53664, + "real_time": 1.3052538210958686e+04, + "cpu_time": 1.2555224753182375e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1380869634697324e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3074507845701832e+04, + "cpu_time": 1.2553521810695126e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1328192340150708e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time_median", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3074344684304739e+04, + "cpu_time": 1.2555224753182378e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1328530025042826e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9065804931666523e+01, + "cpu_time": 7.7941628809930687e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5667061238928000e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4582426471933545e-03, + "cpu_time": 6.2087460383848116e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4576985720430595e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52641, + "real_time": 1.3274283688613998e+04, + "cpu_time": 1.2345705121725705e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1713311182483459e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52641, + "real_time": 1.3294856860828306e+04, + "cpu_time": 1.2344351028590147e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1617812705729401e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52641, + "real_time": 1.3281601897518882e+04, + "cpu_time": 1.2342861868163917e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1679306933076632e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52641, + "real_time": 1.3276541739232571e+04, + "cpu_time": 1.2338799401217234e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1702815092219377e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52641, + "real_time": 1.3263797579979373e+04, + "cpu_time": 1.2330951924577346e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1762100564359939e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3278216353234626e+04, + "cpu_time": 1.2340533868854871e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1695069295573771e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time_median", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3276541739232573e+04, + "cpu_time": 1.2342861868163917e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1702815092219377e+08, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1341732761356489e+01, + "cpu_time": 5.9488271875690133e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2685357404121314e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.5416086465510844e-04, + "cpu_time": 4.8205590218286320e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.5396382572668821e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51505, + "real_time": 1.3600124474675687e+04, + "cpu_time": 1.2306036266103754e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2046948563234158e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51505, + "real_time": 1.3645131789367178e+04, + "cpu_time": 1.2325709460363193e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2007212720925901e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51505, + "real_time": 1.3544414927053445e+04, + "cpu_time": 1.2276043203348753e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2096498880342777e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51505, + "real_time": 1.3635508397656658e+04, + "cpu_time": 1.2325679160301852e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2015686927240415e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51505, + "real_time": 1.3552061037480331e+04, + "cpu_time": 1.2300629439374726e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2089674002122262e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3595448125246661e+04, + "cpu_time": 1.2306819505898459e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2051204218773105e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time_median", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3600124474675689e+04, + "cpu_time": 1.2306036266103754e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2046948563234158e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6319506832165494e+01, + "cpu_time": 2.0607208269044843e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1068147882464826e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.4069863976127763e-03, + "cpu_time": 1.6744544160388593e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4078044929725573e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49826, + "real_time": 1.4002569195946207e+04, + "cpu_time": 1.2135639595679264e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3401419797651458e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49826, + "real_time": 1.4026939635591980e+04, + "cpu_time": 1.2134978375352626e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3360762112966127e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49826, + "real_time": 1.3979937139422416e+04, + "cpu_time": 1.2155145693936088e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3439304249513826e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49826, + "real_time": 1.4047054418359110e+04, + "cpu_time": 1.2152083862013419e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3327310498045149e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49826, + "real_time": 1.4000874655083315e+04, + "cpu_time": 1.2136070666167750e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3404252096566610e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4011475008880607e+04, + "cpu_time": 1.2142783638629830e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3386609750948634e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time_median", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4002569195946207e+04, + "cpu_time": 1.2136070666167750e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3401419797651458e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5939194632590301e+01, + "cpu_time": 9.9541203787659889e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3277597570290333e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8512822251868411e-03, + "cpu_time": 8.1975605223656885e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8505289151000118e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 47094, + "real_time": 1.4844343789965189e+04, + "cpu_time": 1.1997751581656671e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4148802350092764e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 47094, + "real_time": 1.4887003279279079e+04, + "cpu_time": 1.2006135627149208e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4022291639592934e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 47094, + "real_time": 1.4864031905028358e+04, + "cpu_time": 1.2027943128279381e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4090325167984743e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 47094, + "real_time": 1.4852138017903466e+04, + "cpu_time": 1.1998438049181626e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4125633576122046e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 47094, + "real_time": 1.4876955553093369e+04, + "cpu_time": 1.2016866904565871e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4052023793519430e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4864894509053895e+04, + "cpu_time": 1.2009427058166552e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4087815305462389e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time_median", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4864031905028358e+04, + "cpu_time": 1.2006135627149208e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.4090325167984743e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7464755304406271e+01, + "cpu_time": 1.2904502657144736e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1793637964831162e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1748993774405099e-03, + "cpu_time": 1.0745310825106785e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1747834998395586e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 34324, + "real_time": 2.0367021250927894e+04, + "cpu_time": 1.0211278632394733e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4355017056816072e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 34324, + "real_time": 2.0362978027683643e+04, + "cpu_time": 1.0208508512401464e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4367795232016897e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 34324, + "real_time": 2.0386002412190206e+04, + "cpu_time": 1.0210037247887630e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4295096875698853e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 34324, + "real_time": 2.0364474790194439e+04, + "cpu_time": 1.0201987599393852e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4363064282468796e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 34324, + "real_time": 2.0398689616960666e+04, + "cpu_time": 1.0206600000397123e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4255107784481926e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0375833219591372e+04, + "cpu_time": 1.0207682398494959e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4327216246296511e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time_median", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0367021250927894e+04, + "cpu_time": 1.0208508512401464e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.4355017056816072e+09, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5793575016780860e+01, + "cpu_time": 3.6319072380699054e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9839505101876762e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.7511308845987855e-04, + "cpu_time": 3.5580135590870278e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.7478100266380103e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 32434, + "real_time": 2.1616852509860073e+04, + "cpu_time": 1.0479737523315898e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2126834833166786e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 32434, + "real_time": 2.1568266150221654e+04, + "cpu_time": 1.0469047823218400e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2154152687758165e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 32434, + "real_time": 2.1611628625787325e+04, + "cpu_time": 1.0447118083426824e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2129766087466715e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 32434, + "real_time": 2.1618416912420933e+04, + "cpu_time": 1.0444433883155020e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2125957282717787e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 32434, + "real_time": 2.1599212370348130e+04, + "cpu_time": 1.0455401995353409e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2136738854416609e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1602875313727625e+04, + "cpu_time": 1.0459147861693910e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2134689949105213e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time_median", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1611628625787325e+04, + "cpu_time": 1.0455401995353409e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2129766087466715e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0764965239564212e+01, + "cpu_time": 1.4976578782181568e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1674790478770915e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.6121303011775653e-04, + "cpu_time": 1.4319119473425283e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6210043501208616e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 25968, + "real_time": 2.6898262761892776e+04, + "cpu_time": 9.0875978183129597e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9491519011509087e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 25968, + "real_time": 2.6920672822369499e+04, + "cpu_time": 9.0758638457929158e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9475293335326576e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 25968, + "real_time": 2.6895421702741161e+04, + "cpu_time": 9.0785934415496898e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9493577970058929e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 25968, + "real_time": 2.6941948338499755e+04, + "cpu_time": 9.0833310670634592e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9459914086866467e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 25968, + "real_time": 2.6895029128081045e+04, + "cpu_time": 9.0659444108459890e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9493862509060902e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6910266950716847e+04, + "cpu_time": 9.0782661167130017e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9482833382564392e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time_median", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6898262761892773e+04, + "cpu_time": 9.0785934415496886e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9491519011509087e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0668004689578815e+01, + "cpu_time": 8.2224822813998617e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4955553897941727e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.6803417548513956e-04, + "cpu_time": 9.0573267799038727e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6762725442828942e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 17130, + "real_time": 4.0869549459153764e+04, + "cpu_time": 7.2001741794104421e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5656656701047752e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 17130, + "real_time": 4.0856400017248860e+04, + "cpu_time": 7.2019090070886600e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5664914176415676e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 17130, + "real_time": 4.0842285711369288e+04, + "cpu_time": 7.1922786974604797e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5673783475544006e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 17130, + "real_time": 4.0844329318036900e+04, + "cpu_time": 7.1992890952654910e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5672498912522175e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 17130, + "real_time": 4.0856566940374061e+04, + "cpu_time": 7.1923007990272748e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5664809320134220e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0853826289236575e+04, + "cpu_time": 7.1971903556504703e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5666532517132767e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time_median", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0856400017248860e+04, + "cpu_time": 7.1992890952654910e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5664914176415676e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1008300472248621e+01, + "cpu_time": 4.5718085512091574e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9154618175332295e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6945580064672886e-04, + "cpu_time": 6.3522129126678689e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6943498553679045e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10180, + "real_time": 6.8738485796921173e+04, + "cpu_time": 5.7150407440605950e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0509138740643200e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10180, + "real_time": 6.8711042527966143e+04, + "cpu_time": 5.7124072919645059e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0521324125542648e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10180, + "real_time": 6.8747409389419248e+04, + "cpu_time": 5.7179166395331180e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0505178575103775e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10180, + "real_time": 6.8705777898051165e+04, + "cpu_time": 5.7194312885786843e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0523662844074802e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10180, + "real_time": 6.8721432943513006e+04, + "cpu_time": 5.7203001064727938e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0516709419080318e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8724829711174156e+04, + "cpu_time": 5.7170192141219401e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0515202740888947e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time_median", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8721432943513006e+04, + "cpu_time": 5.7179166395331180e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0516709419080318e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7754763978391598e+01, + "cpu_time": 3.2623572745201160e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8830637257147683e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5834569620628402e-04, + "cpu_time": 5.7063955049540128e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5833233987175287e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5614, + "real_time": 1.2463504604171675e+05, + "cpu_time": 4.8108513900277966e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3652685446083271e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5614, + "real_time": 1.2465379725501595e+05, + "cpu_time": 4.8142309735766834e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3647623196101437e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5614, + "real_time": 1.2462139276816751e+05, + "cpu_time": 4.8115753124355286e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3656372367805580e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5614, + "real_time": 1.2466229027271368e+05, + "cpu_time": 4.8137700962917656e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3645330844030365e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5614, + "real_time": 1.2462908839067862e+05, + "cpu_time": 4.8107027296357936e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3654294147221771e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2464032294565851e+05, + "cpu_time": 4.8122261003935134e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3651261200248489e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time_median", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2463504604171675e+05, + "cpu_time": 4.8115753124355286e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3652685446083271e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7150313556686569e+01, + "cpu_time": 1.6611330326008768e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6302125960694291e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3759843645594431e-04, + "cpu_time": 3.4519014650309958e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3759402860167506e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2961, + "real_time": 2.3629125720149485e+05, + "cpu_time": 4.2674041745253813e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5501135756566330e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2961, + "real_time": 2.3625559600085687e+05, + "cpu_time": 4.2669167139655399e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5506494415351646e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2961, + "real_time": 2.3628412691046877e+05, + "cpu_time": 4.2673116227870077e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5502207066065659e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2961, + "real_time": 2.3625473158957096e+05, + "cpu_time": 4.2652958777075738e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5506624326885223e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2961, + "real_time": 2.3626677973499589e+05, + "cpu_time": 4.2675281566219002e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5504813708507484e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3627049828747744e+05, + "cpu_time": 4.2668913091214806e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5504255054675262e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time_median", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3626677973499586e+05, + "cpu_time": 4.2673116227870071e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5504813708507484e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6591914071857552e+01, + "cpu_time": 9.2077630582489975e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4932188813660145e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.0224231091558741e-05, + "cpu_time": 2.1579558491600771e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.0223100795286311e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1522, + "real_time": 4.5980910883262468e+05, + "cpu_time": 3.9811065162409520e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6487350245397339e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1522, + "real_time": 4.5976306284403329e+05, + "cpu_time": 3.9805535495673776e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6491004510493660e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1522, + "real_time": 4.5980311382599251e+05, + "cpu_time": 3.9802589698367858e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6487825974900108e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1522, + "real_time": 4.5988195740300231e+05, + "cpu_time": 3.9815024852832133e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6481570389807320e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1522, + "real_time": 4.5972994752286392e+05, + "cpu_time": 3.9800189367137903e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6493633034784218e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5979743808570335e+05, + "cpu_time": 3.9806880915284240e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6488276831076530e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time_median", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5980311382599251e+05, + "cpu_time": 3.9805535495673782e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6487825974900108e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.7115166051968984e+01, + "cpu_time": 6.0998597533747707e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5323101013059551e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2421810414986061e-04, + "cpu_time": 1.5323631525806562e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2421277448338840e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 772, + "real_time": 9.0632641784800938e+05, + "cpu_time": 3.8421922583939034e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7022458287900269e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 772, + "real_time": 9.0638842551915906e+05, + "cpu_time": 3.8420941178110582e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7019925514583626e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 772, + "real_time": 9.0635626540951175e+05, + "cpu_time": 3.8418814740274489e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7021239087302345e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 772, + "real_time": 9.0629365228821035e+05, + "cpu_time": 3.8436063413210398e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7023796774126976e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 772, + "real_time": 9.0627708858955128e+05, + "cpu_time": 3.8429370397796845e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7024473444673660e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0632836993088829e+05, + "cpu_time": 3.8425422462666261e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7022378621717369e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time_median", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0632641784800927e+05, + "cpu_time": 3.8421922583939028e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7022458287900269e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5342397732877217e+01, + "cpu_time": 7.1555761243529792e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8521652190665929e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.0028664264735288e-05, + "cpu_time": 1.8621984263947290e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0028260960523769e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 389, + "real_time": 1.7994209213642008e+06, + "cpu_time": 3.7888818759386706e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7294700313433365e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 389, + "real_time": 1.7998568919561517e+06, + "cpu_time": 3.7894564247051483e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7285666599339233e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 389, + "real_time": 1.7995336227679197e+06, + "cpu_time": 3.7892208044207621e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7292364616548660e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 389, + "real_time": 1.7993629216946720e+06, + "cpu_time": 3.7887078350649917e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7295902450182579e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 389, + "real_time": 1.7998914571316163e+06, + "cpu_time": 3.7897391978664982e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7284950564156548e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7996131629829120e+06, + "cpu_time": 3.7892012275992143e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7290716908732079e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time_median", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7995336227679201e+06, + "cpu_time": 3.7892208044207615e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7292364616548660e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4635119020380748e+02, + "cpu_time": 4.1843234185226363e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1046216408897536e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3689119154667277e-04, + "cpu_time": 1.1042758531918260e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3688719509960518e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 195, + "real_time": 3.5856968329216423e+06, + "cpu_time": 3.7907158444094980e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7431421074892929e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 195, + "real_time": 3.5861456945825079e+06, + "cpu_time": 3.7915317579478520e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7426735952964500e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 195, + "real_time": 3.5855097839465509e+06, + "cpu_time": 3.7905157351797360e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7433373798318657e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 195, + "real_time": 3.5857165733782146e+06, + "cpu_time": 3.7912797049226028e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7431215003574394e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 195, + "real_time": 3.5861760819665133e+06, + "cpu_time": 3.7916632601035392e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7426418818342140e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5858489933590861e+06, + "cpu_time": 3.7911412605126452e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7429832929618523e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time_median", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5857165733782141e+06, + "cpu_time": 3.7912797049226028e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7431215003574394e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9612895799918545e+02, + "cpu_time": 5.0408320134313813e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0910154412813922e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.2582662724395208e-05, + "cpu_time": 1.3296344470028400e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2581598670066389e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 98, + "real_time": 7.1579838760805372e+06, + "cpu_time": 3.8384357986736095e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7501545218202682e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 98, + "real_time": 7.1584247860449310e+06, + "cpu_time": 3.8376228932662457e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7499235379731087e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 98, + "real_time": 7.1579904381033713e+06, + "cpu_time": 3.8376255834688139e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7501510839001122e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 98, + "real_time": 7.1581635025462937e+06, + "cpu_time": 3.8375226745910257e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7500604156989769e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 98, + "real_time": 7.1586190095665501e+06, + "cpu_time": 3.8377726826530653e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7498217972107666e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1582363224683357e+06, + "cpu_time": 3.8377959265305519e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7500222713206474e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time_median", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1581635025462937e+06, + "cpu_time": 3.8376255834688145e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7500604156989769e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7904254699357961e+02, + "cpu_time": 3.6864635474672847e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4618137993602331e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.8982024960215176e-05, + "cpu_time": 9.6056789314483568e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8981469804589327e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49, + "real_time": 1.4304209253465643e+07, + "cpu_time": 3.9326837924487078e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7532372638489349e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49, + "real_time": 1.4304042546724787e+07, + "cpu_time": 3.9328243820402199e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7532810060253067e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49, + "real_time": 1.4303174154946998e+07, + "cpu_time": 3.9326460344904613e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7535088798056335e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49, + "real_time": 1.4304150333057862e+07, + "cpu_time": 3.9329470644904125e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7532527238563408e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49, + "real_time": 1.4303317730676156e+07, + "cpu_time": 3.9325613291827476e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7534712023391563e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4303778803774288e+07, + "cpu_time": 3.9327325205305105e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7533502151750748e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time_median", + "family_index": 4, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4304042546724787e+07, + "cpu_time": 3.9326837924487078e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7532810060253067e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9271436784810567e+02, + "cpu_time": 1.5299376394035911e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2929116466642260e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.4446447656062381e-05, + "cpu_time": 3.8902661989257493e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4446869397821918e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 24, + "real_time": 2.8600150253623724e+07, + "cpu_time": 4.0750322320838946e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7543223181631844e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 24, + "real_time": 2.8596136951819062e+07, + "cpu_time": 4.0749921487507850e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7548492155046028e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 24, + "real_time": 2.8596603622039158e+07, + "cpu_time": 4.0750283487492800e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7547879398254005e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 24, + "real_time": 2.8596450341865420e+07, + "cpu_time": 4.0747829254176325e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7548080659089142e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 24, + "real_time": 2.8596263689299423e+07, + "cpu_time": 4.0747816233330053e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7548325741652351e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8597120971729361e+07, + "cpu_time": 4.0749234556669199e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7547200227134674e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time_median", + "family_index": 4, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8596450341865420e+07, + "cpu_time": 4.0749921487507856e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7548080659089142e+10, + "dst_id": 4.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7027311058120631e+03, + "cpu_time": 1.2982560378771966e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2354633282252699e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.9542046470179816e-05, + "cpu_time": 3.1859642322157886e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9537417296156786e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57477, + "real_time": 1.2198978392134453e+04, + "cpu_time": 1.1949995928154285e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0985363837111257e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57477, + "real_time": 1.2226356702697263e+04, + "cpu_time": 1.1978124809650681e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0938371603662085e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57477, + "real_time": 1.2203329057869967e+04, + "cpu_time": 1.1898432553986824e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0977882247213908e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57477, + "real_time": 1.2174152770084222e+04, + "cpu_time": 1.1930819748684738e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1028157345706526e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57477, + "real_time": 1.2182130166748550e+04, + "cpu_time": 1.1950896404537573e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1014387179900512e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2196989417906891e+04, + "cpu_time": 1.1941653889002819e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0988832442718860e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time_median", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2198978392134453e+04, + "cpu_time": 1.1949995928154283e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0985363837111257e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0303289746200164e+01, + "cpu_time": 2.9455916894780891e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4920362510870720e+04, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6646148529400287e-03, + "cpu_time": 2.4666530422479517e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6637591731780575e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56729, + "real_time": 1.2305945836356312e+04, + "cpu_time": 1.1780168721420159e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1605903910885319e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56729, + "real_time": 1.2291753889428175e+04, + "cpu_time": 1.1813030573162589e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1653941708055034e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56729, + "real_time": 1.2331564198671300e+04, + "cpu_time": 1.1838180482328448e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1519469205307059e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56729, + "real_time": 1.2334461285801919e+04, + "cpu_time": 1.1863831200262749e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1509717217188746e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56729, + "real_time": 1.2290935355852042e+04, + "cpu_time": 1.1786188211016092e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1656715715799704e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2310932113221950e+04, + "cpu_time": 1.1816279837638009e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1589149551447183e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time_median", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2305945836356310e+04, + "cpu_time": 1.1813030573162587e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1605903910885319e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1046629435573887e+01, + "cpu_time": 3.5216639362579561e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1078768001598059e+04, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7095886194490337e-03, + "cpu_time": 2.9803491324236551e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7090700042728987e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56595, + "real_time": 1.2406941944871231e+04, + "cpu_time": 1.1852753194342685e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2534439554083690e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56595, + "real_time": 1.2444398606986279e+04, + "cpu_time": 1.1883549563621712e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2286017375329569e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56595, + "real_time": 1.2382533652841226e+04, + "cpu_time": 1.1781322533606160e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2697130386157975e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56595, + "real_time": 1.2386013261711139e+04, + "cpu_time": 1.1822065453365545e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2673898240161702e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56595, + "real_time": 1.2395178755884221e+04, + "cpu_time": 1.1850850027529619e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2612765831544638e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2403013244458822e+04, + "cpu_time": 1.1838108154493146e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2560850277455524e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time_median", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2395178755884222e+04, + "cpu_time": 1.1850850027529619e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2612765831544638e+07, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4989874028778353e+01, + "cpu_time": 3.8482883153837323e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6604064641677350e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0148228125083109e-03, + "cpu_time": 3.2507629303277796e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0111305280744352e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55869, + "real_time": 1.2489207323259856e+04, + "cpu_time": 1.1670144746144738e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6398158401821163e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55869, + "real_time": 1.2468664995042232e+04, + "cpu_time": 1.1702678690555153e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6425174634287807e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55869, + "real_time": 1.2481334523702857e+04, + "cpu_time": 1.1724461480077937e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6408501800113735e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55869, + "real_time": 1.2518957053903572e+04, + "cpu_time": 1.1742233016058948e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6359190235910326e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55869, + "real_time": 1.2481119389795449e+04, + "cpu_time": 1.1658256753693824e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6408784629321331e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2487856657140794e+04, + "cpu_time": 1.1699554937306120e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6399961940290877e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time_median", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2481334523702853e+04, + "cpu_time": 1.1702678690555151e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6408501800113735e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8876624881896142e+01, + "cpu_time": 3.5433075745393960e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4759006515811576e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5115984592201524e-03, + "cpu_time": 3.0285832183589547e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5096990228364176e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55399, + "real_time": 1.2638686060726401e+04, + "cpu_time": 1.1681691084118071e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2408432176569033e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55399, + "real_time": 1.2710641545918967e+04, + "cpu_time": 1.1705109153354540e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2224966656502962e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55399, + "real_time": 1.2621833489611387e+04, + "cpu_time": 1.1627724259565907e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2451703655980581e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55399, + "real_time": 1.2637591513674603e+04, + "cpu_time": 1.1653279463570275e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2411239084345239e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55399, + "real_time": 1.2640600785368830e+04, + "cpu_time": 1.1678974444916933e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2403523135870367e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2649870679060039e+04, + "cpu_time": 1.1669355681105146e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2379972941853642e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time_median", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2638686060726401e+04, + "cpu_time": 1.1678974444916933e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2408432176569033e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4788600035458039e+01, + "cpu_time": 2.9642088287362647e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8761415028152859e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7501150737489627e-03, + "cpu_time": 2.5401649497545685e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7412442619253025e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54618, + "real_time": 1.2781480106605271e+04, + "cpu_time": 1.1500677831190727e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4092733640186954e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54618, + "real_time": 1.2773146172060644e+04, + "cpu_time": 1.1526851984290340e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4134551422567928e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54618, + "real_time": 1.2749525284998548e+04, + "cpu_time": 1.1553474276799679e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4253372709013236e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54618, + "real_time": 1.2818613636520291e+04, + "cpu_time": 1.1557424361543393e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.3907066959729195e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54618, + "real_time": 1.2746135802780605e+04, + "cpu_time": 1.1493885681170301e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4270459116031802e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2773780200593072e+04, + "cpu_time": 1.1526462826998889e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4131636769505823e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time_median", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2773146172060642e+04, + "cpu_time": 1.1526851984290338e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4134551422567928e+08, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9249962318918051e+01, + "cpu_time": 2.9216959647611640e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4666699611841787e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2898438723378072e-03, + "cpu_time": 2.5347723830050971e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2869679226424650e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53422, + "real_time": 1.3097699670007492e+04, + "cpu_time": 1.1516102059417171e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2509066792482522e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53422, + "real_time": 1.3141534501638123e+04, + "cpu_time": 1.1539198740642653e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2467341616732581e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53422, + "real_time": 1.3101635108725139e+04, + "cpu_time": 1.1470285073134172e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2505309348059118e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53422, + "real_time": 1.3103179955924399e+04, + "cpu_time": 1.1496559713774114e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2503834988996108e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53422, + "real_time": 1.3132552102929827e+04, + "cpu_time": 1.1524599824821696e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2475869024989276e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3115320267844998e+04, + "cpu_time": 1.1509349082357962e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2492284354251919e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time_median", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3103179955924399e+04, + "cpu_time": 1.1516102059417171e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2503834988996108e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0182160535852276e+01, + "cpu_time": 2.6733661879370660e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9211461746363810e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5388233092052768e-03, + "cpu_time": 2.3227779162897401e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5378661901676075e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51563, + "real_time": 1.3541849375179898e+04, + "cpu_time": 1.1416491624496834e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4197581210775127e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51563, + "real_time": 1.3530598470151426e+04, + "cpu_time": 1.1439145903258312e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4217701879400520e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51563, + "real_time": 1.3547310286956601e+04, + "cpu_time": 1.1473846275383112e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4187827181864395e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51563, + "real_time": 1.3560765273471288e+04, + "cpu_time": 1.1479868784852996e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4163828028277669e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51563, + "real_time": 1.3531358413166185e+04, + "cpu_time": 1.1418811684851179e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4216341774020495e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3542376363785083e+04, + "cpu_time": 1.1445632854568486e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4196656014867644e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time_median", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3541849375179900e+04, + "cpu_time": 1.1439145903258309e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4197581210775127e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2478855400178384e+01, + "cpu_time": 2.9911381650419976e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2287544225059878e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.2146718308237568e-04, + "cpu_time": 2.6133444983324752e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.2110018059376835e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 48627, + "real_time": 1.4413517360083553e+04, + "cpu_time": 1.1380836594654038e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5468429643338699e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 48627, + "real_time": 1.4450651741714451e+04, + "cpu_time": 1.1392448395761445e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5351587714772987e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 48627, + "real_time": 1.4392307976238395e+04, + "cpu_time": 1.1345194756647594e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5535434697617292e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 48627, + "real_time": 1.4390981724045207e+04, + "cpu_time": 1.1364168461006713e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5539631177836199e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 48627, + "real_time": 1.4410197455703046e+04, + "cpu_time": 1.1386476280785291e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5478904922335510e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4411531251556929e+04, + "cpu_time": 1.1373824897771015e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5474797631180143e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time_median", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4410197455703048e+04, + "cpu_time": 1.1380836594654038e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5478904922335510e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4124436695971490e+01, + "cpu_time": 1.9163688413682089e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6023593717108639e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6739676218212578e-03, + "cpu_time": 1.6848939196732039e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6717742063129596e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 40705, + "real_time": 1.7191178507388198e+04, + "cpu_time": 1.0668983864299673e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.6243754867456932e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 40705, + "real_time": 1.7111289011021050e+04, + "cpu_time": 1.0663300624777750e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.6599723092502890e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 40705, + "real_time": 1.7026425225034454e+04, + "cpu_time": 1.0660783213979577e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.6981514479786959e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 40705, + "real_time": 1.7175690154160700e+04, + "cpu_time": 1.0687497197629262e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.6312508448604412e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 40705, + "real_time": 1.7009521921486619e+04, + "cpu_time": 1.0631634573857970e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.7058015272274275e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7102820963818205e+04, + "cpu_time": 1.0662439894908845e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.6639103232125092e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time_median", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7111289011021050e+04, + "cpu_time": 1.0663300624777750e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.6599723092502890e+09, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.3261189089698149e+01, + "cpu_time": 2.0144138152621891e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7324524021613456e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.8682722730852990e-03, + "cpu_time": 1.8892615903270329e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8701671141120556e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 36374, + "real_time": 1.9224827358176884e+04, + "cpu_time": 1.0499035476418051e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3635701123136612e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 36374, + "real_time": 1.9277247081422051e+04, + "cpu_time": 1.0521908895926411e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3598622193966406e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 36374, + "real_time": 1.9192667634373604e+04, + "cpu_time": 1.0423166263230475e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3658549452005642e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 36374, + "real_time": 1.9237534214704046e+04, + "cpu_time": 1.0473549773570226e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3626694412823055e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 36374, + "real_time": 1.9220288104725842e+04, + "cpu_time": 1.0514130358546554e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3638921465258610e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9230512878680489e+04, + "cpu_time": 1.0486358153538347e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3631697729438065e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time_median", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9224827358176888e+04, + "cpu_time": 1.0499035476418053e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3635701123136612e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0832282301091869e+01, + "cpu_time": 3.9853277056764639e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1841753301754877e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6033000521412739e-03, + "cpu_time": 3.8004878789417653e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6022768209264901e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27402, + "real_time": 2.5494062332736106e+04, + "cpu_time": 8.7151190686316752e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0565102303322552e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27402, + "real_time": 2.5503797527115694e+04, + "cpu_time": 8.7281680847765231e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0557252285373413e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27402, + "real_time": 2.5488598452206421e+04, + "cpu_time": 8.7508097369996727e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0569510755292828e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27402, + "real_time": 2.5557137056268108e+04, + "cpu_time": 8.7540613998715854e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0514347864774387e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27402, + "real_time": 2.5485894205162411e+04, + "cpu_time": 8.7170659818326521e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0571693336692909e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5505897914697747e+04, + "cpu_time": 8.7330448544224226e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0555581309091217e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time_median", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5494062332736103e+04, + "cpu_time": 8.7281680847765219e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0565102303322552e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9450645387856547e+01, + "cpu_time": 1.8423675550737088e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3702665083113335e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1546602078606159e-03, + "cpu_time": 2.1096508557845459e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1531011809736678e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 17621, + "real_time": 3.9783550985096044e+04, + "cpu_time": 6.9863605545146513e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6357023795910622e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 17621, + "real_time": 3.9782096810373936e+04, + "cpu_time": 6.9841478735694802e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6357987237278149e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 17621, + "real_time": 3.9732819274628651e+04, + "cpu_time": 6.9608570549026096e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6390677005635166e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 17621, + "real_time": 3.9689326619857609e+04, + "cpu_time": 6.9591650233211660e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6419596634712620e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 17621, + "real_time": 3.9774706088061648e+04, + "cpu_time": 6.9828800377279091e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6362884936935570e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9752499955603584e+04, + "cpu_time": 6.9746821088071632e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6377633922094421e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time_median", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9774706088061648e+04, + "cpu_time": 6.9828800377279079e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.6362884936935570e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0961726388659635e+01, + "cpu_time": 1.3463885886967671e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7199693792158764e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0304188776657202e-03, + "cpu_time": 1.9303941996103842e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0311650344565504e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10284, + "real_time": 6.7978140837560219e+04, + "cpu_time": 5.5508890650994420e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0850387700530525e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10284, + "real_time": 6.7987324133975766e+04, + "cpu_time": 5.5463587405387640e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0846220625882465e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10284, + "real_time": 6.8034496310921662e+04, + "cpu_time": 5.5508053622122931e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0824833190737415e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10284, + "real_time": 6.7952286230165148e+04, + "cpu_time": 5.5514107056095600e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0862125711217632e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10284, + "real_time": 6.8047796172090646e+04, + "cpu_time": 5.5527263879084384e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0818808513597874e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8000008736942691e+04, + "cpu_time": 5.5504380522737002e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0840475148393185e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time_median", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7987324133975766e+04, + "cpu_time": 5.5508890650994420e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0846220625882465e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9967404038648255e+01, + "cpu_time": 2.4063274843471832e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8125465325641710e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.8775586622733728e-04, + "cpu_time": 4.3353830124479037e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8771679873375953e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5609, + "real_time": 1.2473712897425565e+05, + "cpu_time": 4.6689316051830834e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3625144610035534e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5609, + "real_time": 1.2470883329267426e+05, + "cpu_time": 4.6680418809138405e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3632773952399609e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5609, + "real_time": 1.2477709664871449e+05, + "cpu_time": 4.6750961376740479e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3614374053022270e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5609, + "real_time": 1.2478964731887078e+05, + "cpu_time": 4.6730786330860418e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3610993300449329e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5609, + "real_time": 1.2471168553810746e+05, + "cpu_time": 4.6705851066117018e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3632004746807545e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2474487835452454e+05, + "cpu_time": 4.6711466726937437e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3623058132542862e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time_median", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2473712897425567e+05, + "cpu_time": 4.6705851066117018e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3625144610035534e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7092253467637093e+01, + "cpu_time": 2.9237115581405832e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9970728198368140e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9734490070382712e-04, + "cpu_time": 6.2590874639664132e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9732788672666611e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2935, + "real_time": 2.3830264284266747e+05, + "cpu_time": 4.1694605180824208e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5201489584563019e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2935, + "real_time": 2.3831855743691910e+05, + "cpu_time": 4.1727424101726609e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5199138876209396e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2935, + "real_time": 2.3830182443998018e+05, + "cpu_time": 4.1708492279420191e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5201610477442207e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2935, + "real_time": 2.3826589627691990e+05, + "cpu_time": 4.1720168237995851e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5206918535460495e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2935, + "real_time": 2.3833578415282446e+05, + "cpu_time": 4.1717334204227459e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5196594711187386e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3830494102986221e+05, + "cpu_time": 4.1713604800838864e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5201150436972496e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time_median", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3830264284266747e+05, + "cpu_time": 4.1717334204227459e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.5201489584563019e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5865963466268546e+01, + "cpu_time": 1.2600433249920269e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8209499521663459e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0854144842522278e-04, + "cpu_time": 3.0207011142002467e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0854616695007567e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1505, + "real_time": 4.6501518805368966e+05, + "cpu_time": 3.9424255680588847e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6078855983652168e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1505, + "real_time": 4.6511533267121529e+05, + "cpu_time": 3.9428855054949760e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6071087795894318e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1505, + "real_time": 4.6496947256900812e+05, + "cpu_time": 3.9418402561451507e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6082403232418701e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1505, + "real_time": 4.6492631236196449e+05, + "cpu_time": 3.9410095234228379e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6085752847083939e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1505, + "real_time": 4.6512798514090281e+05, + "cpu_time": 3.9434428260455287e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6070106585647865e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6503085815935611e+05, + "cpu_time": 3.9423207358334768e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6077641288939400e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time_median", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6501518805368972e+05, + "cpu_time": 3.9424255680588853e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6078855983652168e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8759732299025359e+01, + "cpu_time": 9.4063538425926687e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8860182496418059e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9086847838516838e-04, + "cpu_time": 2.3859940560122888e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9086664215359626e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 761, + "real_time": 9.1871143183787202e+05, + "cpu_time": 3.8204097788441104e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6523363960841034e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 761, + "real_time": 9.1867798828592093e+05, + "cpu_time": 3.8205125314194620e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6524693557321663e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 761, + "real_time": 9.1865067458962451e+05, + "cpu_time": 3.8197312562285244e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6525779524397873e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 761, + "real_time": 9.1875475332931511e+05, + "cpu_time": 3.8204262196457523e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6521641796581673e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 761, + "real_time": 9.1869607503536937e+05, + "cpu_time": 3.8201999273314542e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6523974480579086e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1869818461562041e+05, + "cpu_time": 3.8202559426938605e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6523890663944267e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time_median", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1869607503536926e+05, + "cpu_time": 3.8204097788441104e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.6523974480579086e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8874966358845676e+01, + "cpu_time": 3.1501236941909952e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5454996903267240e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.2315275037917711e-05, + "cpu_time": 8.2458446278070043e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2314760619202424e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 383, + "real_time": 1.8255143786851054e+06, + "cpu_time": 3.7730179863190192e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6761618962616798e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 383, + "real_time": 1.8261000577861310e+06, + "cpu_time": 3.7740947735250950e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6749828528760521e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 383, + "real_time": 1.8257178103565858e+06, + "cpu_time": 3.7732686649601376e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6757522777790497e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 383, + "real_time": 1.8255724486537452e+06, + "cpu_time": 3.7729084369971079e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6760449605540955e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 383, + "real_time": 1.8260570330540442e+06, + "cpu_time": 3.7741580896086007e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6750694411642639e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8257923457071227e+06, + "cpu_time": 3.7734895902819920e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6756022857270287e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time_median", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8257178103565858e+06, + "cpu_time": 3.7732686649601370e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.6757522777790497e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7199356899163109e+02, + "cpu_time": 5.9625823597129318e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4754975488990955e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4897289367608177e-04, + "cpu_time": 1.5801242370109070e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4896871650562843e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 192, + "real_time": 3.6393433947523590e+06, + "cpu_time": 3.7804111715624112e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6879654773311905e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 192, + "real_time": 3.6401034158188850e+06, + "cpu_time": 3.7821681780208868e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6871954630938995e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 192, + "real_time": 3.6395608864647024e+06, + "cpu_time": 3.7814695090109038e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6877450930728279e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 192, + "real_time": 3.6394359000648060e+06, + "cpu_time": 3.7817059424484253e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6878717385188744e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 192, + "real_time": 3.6394450532194846e+06, + "cpu_time": 3.7810912213024276e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6878624635717422e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6395777300640480e+06, + "cpu_time": 3.7813692044690114e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6877280471177078e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time_median", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6394450532194851e+06, + "cpu_time": 3.7814695090109038e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6878624635717422e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0383430402444787e+02, + "cpu_time": 6.6274314374347829e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0782478332453999e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.3480647085699443e-05, + "cpu_time": 1.7526538878039606e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.3472745113385681e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 96, + "real_time": 7.2676061778717367e+06, + "cpu_time": 3.8134034176047039e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6935883622495529e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 96, + "real_time": 7.2676128232463570e+06, + "cpu_time": 3.8133896179163432e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6935849848987007e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 96, + "real_time": 7.2673193208174780e+06, + "cpu_time": 3.8128898265613508e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6937341562941605e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 96, + "real_time": 7.2680111334193498e+06, + "cpu_time": 3.8137744711446685e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6933825646702103e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 96, + "real_time": 7.2677478165132925e+06, + "cpu_time": 3.8136916068758827e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6935163791742859e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2676594543736447e+06, + "cpu_time": 3.8134297880205899e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6935612894573822e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time_median", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2676128232463580e+06, + "cpu_time": 3.8134034176047039e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6935849848987007e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5109246399253215e+02, + "cpu_time": 3.4685118833025386e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2760958058076988e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.4549288607823516e-05, + "cpu_time": 9.0955178831361543e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4549198072063638e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 48, + "real_time": 1.4522910156908134e+07, + "cpu_time": 3.9118450600000662e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6967171606761322e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 48, + "real_time": 1.4523443416692317e+07, + "cpu_time": 3.9120606018744773e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6965814276726891e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 48, + "real_time": 1.4523476847292235e+07, + "cpu_time": 3.9119541447912335e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6965729187642456e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 48, + "real_time": 1.4523150185899189e+07, + "cpu_time": 3.9119035874995708e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6966560637874458e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 48, + "real_time": 1.4524183468893170e+07, + "cpu_time": 3.9123670493739367e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6963930753823837e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4523432815137010e+07, + "cpu_time": 3.9120260887078571e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6965841292565796e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time_median", + "family_index": 5, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4523443416692317e+07, + "cpu_time": 3.9119541447912341e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6965814276726891e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7921372661130243e+02, + "cpu_time": 2.0639719046190574e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2196974859037793e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.2995899296744994e-05, + "cpu_time": 5.2759666163187256e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2995258413043959e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 24, + "real_time": 2.9036594089120626e+07, + "cpu_time": 4.1335909170827752e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6978917730654488e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 24, + "real_time": 2.9032780788838863e+07, + "cpu_time": 4.1329036312508076e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6983774713470818e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 24, + "real_time": 2.9036467662081122e+07, + "cpu_time": 4.1337810091664553e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6979078739739586e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 24, + "real_time": 2.9034207575023174e+07, + "cpu_time": 4.1330872279164052e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6981957273174965e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 24, + "real_time": 2.9034347583850224e+07, + "cpu_time": 4.1330090200006002e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6981778939549767e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9034879539782811e+07, + "cpu_time": 4.1332743610834092e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6981101479317924e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time_median", + "family_index": 5, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9034347583850227e+07, + "cpu_time": 4.1330872279164058e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6981778939549767e+10, + "dst_id": 5.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6279649656087049e+03, + "cpu_time": 3.8722256987485525e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0735085807008371e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.6069286024696998e-05, + "cpu_time": 9.3684216446100348e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6069411070961987e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55405, + "real_time": 1.2602100299515927e+04, + "cpu_time": 1.2745351374899428e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0314074155546397e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55405, + "real_time": 1.2601323040337098e+04, + "cpu_time": 1.2790533848994684e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0315327143073682e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55405, + "real_time": 1.2655921260238440e+04, + "cpu_time": 1.2801656710770497e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0227685897848018e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55405, + "real_time": 1.2611188149881351e+04, + "cpu_time": 1.2760930254796600e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0299435466150627e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55405, + "real_time": 1.2606103382540492e+04, + "cpu_time": 1.2746100549615443e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0307623397295088e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2615327226502663e+04, + "cpu_time": 1.2768914547815332e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0292829211982764e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time_median", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2606103382540492e+04, + "cpu_time": 1.2760930254796603e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0307623397295088e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3028002819241593e+01, + "cpu_time": 2.5879016523604970e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6958309010584817e+04, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8253987713345769e-03, + "cpu_time": 2.0267201590782576e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8212496948804562e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55516, + "real_time": 1.2699489862975559e+04, + "cpu_time": 1.2861755322958682e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0316580077180803e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55516, + "real_time": 1.2633213725007790e+04, + "cpu_time": 1.2806476011040010e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0528088192356162e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55516, + "real_time": 1.2613073830590272e+04, + "cpu_time": 1.2826717446692586e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0592801316857055e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55516, + "real_time": 1.2598924484017485e+04, + "cpu_time": 1.2830369655139730e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0638389463283449e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55516, + "real_time": 1.2696050709841256e+04, + "cpu_time": 1.2875363489191334e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0327501181381285e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2648150522486472e+04, + "cpu_time": 1.2840136385004466e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0480672046211757e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time_median", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2633213725007790e+04, + "cpu_time": 1.2830369655139728e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0528088192356162e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6922385188044402e+01, + "cpu_time": 2.7914176689458610e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5006959721242674e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7098218513942878e-03, + "cpu_time": 2.1739782080553737e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7071913490248115e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54866, + "real_time": 1.2715197097694359e+04, + "cpu_time": 1.2667396453376040e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0533553049341366e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54866, + "real_time": 1.2709954086711585e+04, + "cpu_time": 1.2710827086894693e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0566774121600062e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54866, + "real_time": 1.2757458223762822e+04, + "cpu_time": 1.2732002673606913e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0266772741033554e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54866, + "real_time": 1.2752326115784217e+04, + "cpu_time": 1.2707581491446142e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0299075690398321e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54866, + "real_time": 1.2716833651786756e+04, + "cpu_time": 1.2700158672069912e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0523189029536828e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2730353835147949e+04, + "cpu_time": 1.2703593275478740e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0437872926382035e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time_median", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2716833651786754e+04, + "cpu_time": 1.2707581491446142e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.0523189029536828e+07, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2616922773957754e+01, + "cpu_time": 2.3437865874992954e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4281893051605590e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7766138370414672e-03, + "cpu_time": 1.8449792406558050e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7755184879983830e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54357, + "real_time": 1.2923410422115372e+04, + "cpu_time": 1.2678466298191280e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5847210087016430e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54357, + "real_time": 1.2886242101301425e+04, + "cpu_time": 1.2622209245766923e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5892918850198892e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54357, + "real_time": 1.2903871045200258e+04, + "cpu_time": 1.2634644013773551e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5871206344407609e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54357, + "real_time": 1.2860259349882190e+04, + "cpu_time": 1.2616543927548225e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5925028759383157e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54357, + "real_time": 1.2949906747374924e+04, + "cpu_time": 1.2699753747641759e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5814785696546814e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2904737933174836e+04, + "cpu_time": 1.2650323446584349e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5870229947510582e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time_median", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2903871045200260e+04, + "cpu_time": 1.2634644013773551e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5871206344407609e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4314994108338290e+01, + "cpu_time": 3.6785471203001318e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2197832825869153e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6591004238933880e-03, + "cpu_time": 2.9078680366021454e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6589301456522588e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53729, + "real_time": 1.3000311814372033e+04, + "cpu_time": 1.2486518568391099e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1506936591104013e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53729, + "real_time": 1.2975615233430643e+04, + "cpu_time": 1.2524227704515820e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1566903968044472e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53729, + "real_time": 1.3016973751299271e+04, + "cpu_time": 1.2557323572106016e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1466607202700734e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53729, + "real_time": 1.3027182408408818e+04, + "cpu_time": 1.2528629156908286e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1441948623948824e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53729, + "real_time": 1.3014142130961967e+04, + "cpu_time": 1.2553911951313832e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1473453715056634e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3006845067694547e+04, + "cpu_time": 1.2530122190647008e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1491170020170933e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time_median", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3014142130961965e+04, + "cpu_time": 1.2528629156908283e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1473453715056634e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9921287220831275e+01, + "cpu_time": 2.8478579038101109e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8280847482205613e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5316002548773581e-03, + "cpu_time": 2.2728093632924558e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5331550860536603e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53136, + "real_time": 1.3185065650358016e+04, + "cpu_time": 1.2449829398002958e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2130900347679055e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53136, + "real_time": 1.3177853442595295e+04, + "cpu_time": 1.2415651274097798e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2164904441270196e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53136, + "real_time": 1.3158260471634558e+04, + "cpu_time": 1.2430274154226015e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2257469501075816e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53136, + "real_time": 1.3133914777783411e+04, + "cpu_time": 1.2429448691858056e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2372873119727600e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53136, + "real_time": 1.3190657319721693e+04, + "cpu_time": 1.2446766920432801e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2104562353779972e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3169150332418596e+04, + "cpu_time": 1.2434394087723527e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2206141952706528e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time_median", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3177853442595293e+04, + "cpu_time": 1.2430274154226012e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2164904441270196e+08, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3193346854030203e+01, + "cpu_time": 1.4000532556251648e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0967457052206770e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7611877963709601e-03, + "cpu_time": 1.1259521338538216e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7630826648186925e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52127, + "real_time": 1.3450513310341978e+04, + "cpu_time": 1.2355814793672571e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2180947761601403e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52127, + "real_time": 1.3419082019573470e+04, + "cpu_time": 1.2381511946899669e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2209478991261709e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52127, + "real_time": 1.3446321115382889e+04, + "cpu_time": 1.2398121336786406e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2184745447776299e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52127, + "real_time": 1.3410321655167723e+04, + "cpu_time": 1.2357290320822456e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2217454898769231e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52127, + "real_time": 1.3441685862593171e+04, + "cpu_time": 1.2351666848655016e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2188947255191393e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3433584792611848e+04, + "cpu_time": 1.2368881049367228e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2196314870920007e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time_median", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3441685862593173e+04, + "cpu_time": 1.2357290320822458e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2188947255191393e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7789884445615986e+01, + "cpu_time": 2.0099345218055774e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6159178537537111e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3242842264560681e-03, + "cpu_time": 1.6249930076806765e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3249230368810715e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51143, + "real_time": 1.3720779022844319e+04, + "cpu_time": 1.2436417876111162e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3882025900601664e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51143, + "real_time": 1.3697693528459540e+04, + "cpu_time": 1.2398407582199206e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3922275623935013e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51143, + "real_time": 1.3710082090848791e+04, + "cpu_time": 1.2412935170354517e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3900659224988880e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51143, + "real_time": 1.3672168706882772e+04, + "cpu_time": 1.2415564048012195e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3966936557405195e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51143, + "real_time": 1.3705283066520040e+04, + "cpu_time": 1.2442915525021627e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3909028249148192e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3701201283111091e+04, + "cpu_time": 1.2421248040339742e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3916185111215792e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time_median", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3705283066520040e+04, + "cpu_time": 1.2415564048012195e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3909028249148192e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8262067379645494e+01, + "cpu_time": 1.8184484554300751e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1905834587172298e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3328807454391896e-03, + "cpu_time": 1.4639820809667511e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3340687253758402e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49168, + "real_time": 1.4242481205645125e+04, + "cpu_time": 1.2379197544292002e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6014454260978251e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49168, + "real_time": 1.4235748027450934e+04, + "cpu_time": 1.2418912556705341e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6036218029165926e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49168, + "real_time": 1.4230986962566729e+04, + "cpu_time": 1.2439043715107996e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6051619731214905e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49168, + "real_time": 1.4253484174934834e+04, + "cpu_time": 1.2434762913359928e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.5978933428253956e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49168, + "real_time": 1.4207579912919526e+04, + "cpu_time": 1.2396657383007510e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6127489974844675e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4234056056703432e+04, + "cpu_time": 1.2413714822494557e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6041743084891539e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time_median", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4235748027450938e+04, + "cpu_time": 1.2418912556705339e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6036218029165926e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7041883742075701e+01, + "cpu_time": 2.5465614597227229e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5159623818876790e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1972612496527256e-03, + "cpu_time": 2.0514096675623380e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1980350899655069e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 44616, + "real_time": 1.5710142185845709e+04, + "cpu_time": 1.2188088731461487e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3431453674614925e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 44616, + "real_time": 1.5711546145945402e+04, + "cpu_time": 1.2164664213551292e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3423998365574656e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 44616, + "real_time": 1.5721188742580378e+04, + "cpu_time": 1.2174558259551857e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3372830226886940e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 44616, + "real_time": 1.5670543916529166e+04, + "cpu_time": 1.2192973376511409e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3642278594903316e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 44616, + "real_time": 1.5703309931671583e+04, + "cpu_time": 1.2213655661348448e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3467753340105972e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5703346184514448e+04, + "cpu_time": 1.2186788048484898e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3467662840417166e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time_median", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5710142185845709e+04, + "cpu_time": 1.2188088731461484e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3431453674614925e+09, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9415019824054035e+01, + "cpu_time": 1.8713968642974521e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0332180861227700e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2363619572495817e-03, + "cpu_time": 1.5355948235516500e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2378663196766300e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 32445, + "real_time": 2.1603370358709923e+04, + "cpu_time": 1.0298405686970378e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2134402903216919e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 32445, + "real_time": 2.1527525443072300e+04, + "cpu_time": 1.0291317786380367e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2177154345640766e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 32445, + "real_time": 2.1683727972333458e+04, + "cpu_time": 1.0364387655204451e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2089434083220045e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 32445, + "real_time": 2.1550559305553019e+04, + "cpu_time": 1.0298196382840052e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2164139050091953e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 32445, + "real_time": 2.1635024552862327e+04, + "cpu_time": 1.0313055371274347e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2116649064089838e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1600041526506207e+04, + "cpu_time": 1.0313072576533921e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2136355889251904e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time_median", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1603370358709923e+04, + "cpu_time": 1.0298405686970379e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.2134402903216919e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3137522557474632e+01, + "cpu_time": 2.9761246940374346e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5462117336249396e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9230278321454267e-03, + "cpu_time": 2.8857788713804134e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9219740801812721e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 29854, + "real_time": 2.3462678050325369e+04, + "cpu_time": 9.7870336919120884e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.2345616253841465e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 29854, + "real_time": 2.3490267683239068e+04, + "cpu_time": 9.7673669054679132e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.2319371029309021e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 29854, + "real_time": 2.3452699460872431e+04, + "cpu_time": 9.7730176247023439e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.2355123804605164e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 29854, + "real_time": 2.3429379816443692e+04, + "cpu_time": 9.7777465552571678e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.2377374224478333e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 29854, + "real_time": 2.3449076517456178e+04, + "cpu_time": 9.7865607019362235e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.2358577729477093e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3456820305667348e+04, + "cpu_time": 9.7783450958551478e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.2351212608342216e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time_median", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3452699460872431e+04, + "cpu_time": 9.7777465552571678e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.2355123804605164e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2265266139595891e+01, + "cpu_time": 8.5476421980273724e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1208358985539641e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.4920222986132638e-04, + "cpu_time": 8.7413996072306264e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.4886838388463877e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 20797, + "real_time": 3.3710272120293252e+04, + "cpu_time": 7.8897472421686435e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1105533537617683e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 20797, + "real_time": 3.3627061452325870e+04, + "cpu_time": 7.8916187532882357e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1182504646937370e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 20797, + "real_time": 3.3710294748986737e+04, + "cpu_time": 7.9248993475925934e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1105512657421009e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 20797, + "real_time": 3.3687222476778785e+04, + "cpu_time": 7.8950438894109082e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1126816724733025e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 20797, + "real_time": 3.3689315379379943e+04, + "cpu_time": 7.9005477525791979e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1124883013853016e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3684833235552920e+04, + "cpu_time": 7.9003713970079160e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1129050116112423e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time_median", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3689315379379943e+04, + "cpu_time": 7.8950438894109082e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1124883013853016e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4127710689239862e+01, + "cpu_time": 1.4313588268610931e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1570480949930426e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0131476813493470e-03, + "cpu_time": 1.8117614412446323e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0141806715004617e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 12882, + "real_time": 5.4304417998284633e+04, + "cpu_time": 6.2458962600316572e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.8618441690439346e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 12882, + "real_time": 5.4306900533898486e+04, + "cpu_time": 6.2471629476648009e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.8616676322578072e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 12882, + "real_time": 5.4343317726050991e+04, + "cpu_time": 6.2457117527673292e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.8590798054912857e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 12882, + "real_time": 5.4331181962501498e+04, + "cpu_time": 6.2414883526030731e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.8599417944697403e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 12882, + "real_time": 5.4315112012645804e+04, + "cpu_time": 6.2496545713012290e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.8610838168054131e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4320186046676281e+04, + "cpu_time": 6.2459827768736172e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.8607234436136368e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time_median", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4315112012645797e+04, + "cpu_time": 6.2458962600316572e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.8610838168054131e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6634803679175469e+01, + "cpu_time": 2.9646106911196827e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1821493314156212e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.0623613227100333e-04, + "cpu_time": 4.7464278993795072e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0619891548334515e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 7321, + "real_time": 9.5563390209745878e+04, + "cpu_time": 5.1808055986795276e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.3890280480780296e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 7321, + "real_time": 9.5589503674297099e+04, + "cpu_time": 5.1781222588638276e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.3878290385221443e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 7321, + "real_time": 9.5543855186836139e+04, + "cpu_time": 5.1741666301018202e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.3899254345536224e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 7321, + "real_time": 9.5581287550853158e+04, + "cpu_time": 5.1788951873755848e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.3882062142848396e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 7321, + "real_time": 9.5576016983712092e+04, + "cpu_time": 5.1797862103458500e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.3884482031876114e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5570810721088870e+04, + "cpu_time": 5.1783551770733225e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.3886873877252495e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time_median", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5576016983712107e+04, + "cpu_time": 5.1788951873755848e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.3884482031876114e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7804216032514788e+01, + "cpu_time": 2.5466747946729013e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1766587351069022e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8629344983243996e-04, + "cpu_time": 4.9179222119565360e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8631217064984525e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3932, + "real_time": 1.7792729774471288e+05, + "cpu_time": 4.4753143107799965e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7146267640369804e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3932, + "real_time": 1.7799863105181852e+05, + "cpu_time": 4.4740784064263821e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7127373679396049e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3932, + "real_time": 1.7795574175388008e+05, + "cpu_time": 4.4762442771526605e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7138731896618317e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3932, + "real_time": 1.7799053369188256e+05, + "cpu_time": 4.4778746509487009e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7129517654694077e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3932, + "real_time": 1.7796037886674269e+05, + "cpu_time": 4.4768216838489789e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7137503602874535e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7796651662180736e+05, + "cpu_time": 4.4760666658313447e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7135878894790558e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time_median", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7796037886674266e+05, + "cpu_time": 4.4762442771526605e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7137503602874535e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8723015295683599e+01, + "cpu_time": 1.4482421903918716e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6077164121278860e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6139561441617825e-04, + "cpu_time": 3.2355241744883354e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6139969361998442e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2042, + "real_time": 3.4275465037107712e+05, + "cpu_time": 4.1367039501461190e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.8948179059967384e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2042, + "real_time": 3.4268958158902806e+05, + "cpu_time": 4.1364845732365394e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.8957473180845482e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2042, + "real_time": 3.4271207967662252e+05, + "cpu_time": 4.1359650300972635e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.8954259259932434e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2042, + "real_time": 3.4262790166440874e+05, + "cpu_time": 4.1358450139823270e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.8966286512277855e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2042, + "real_time": 3.4265472188461514e+05, + "cpu_time": 4.1347734901768911e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.8962453830271523e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4268778703715035e+05, + "cpu_time": 4.1359544115278286e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.8957730368658936e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time_median", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4268958158902812e+05, + "cpu_time": 4.1359650300972641e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.8957473180845482e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9354889066829990e+01, + "cpu_time": 7.4992255796247133e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.0509028359437771e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4402290053447248e-04, + "cpu_time": 1.8131789747785171e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4402021463922934e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1041, + "real_time": 6.7174735407038126e+05, + "cpu_time": 3.9498428867242044e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9950970103090843e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1041, + "real_time": 6.7171074709909584e+05, + "cpu_time": 3.9496643331029183e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9953692336933525e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1041, + "real_time": 6.7172705497338611e+05, + "cpu_time": 3.9503677203658336e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9952479584627464e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1041, + "real_time": 6.7177411801996757e+05, + "cpu_time": 3.9511850644950229e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9948980021589104e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1041, + "real_time": 6.7174422600969498e+05, + "cpu_time": 3.9501153415281516e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9951202706006920e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7174070003450534e+05, + "cpu_time": 3.9502350692432272e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9951464950449577e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time_median", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7174422600969509e+05, + "cpu_time": 3.9501153415281522e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9951202706006920e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3753086820658261e+01, + "cpu_time": 5.9455211954934952e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7662990811750994e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.5360499697931261e-05, + "cpu_time": 1.5051056687197397e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5360305907488746e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 526, + "real_time": 1.3300502704219192e+06, + "cpu_time": 3.8707004535750175e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0455885384476257e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 526, + "real_time": 1.3304338531529694e+06, + "cpu_time": 3.8717980703994554e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0441338245385147e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 526, + "real_time": 1.3300840341850775e+06, + "cpu_time": 3.8714925148851603e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0454604577760071e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 526, + "real_time": 1.3300572865476173e+06, + "cpu_time": 3.8704395082313198e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0455619226892181e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 526, + "real_time": 1.3305384981332649e+06, + "cpu_time": 3.8723423285362905e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0437371105122627e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3302327884881697e+06, + "cpu_time": 3.8713545751254487e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0448963707927254e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time_median", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3300840341850775e+06, + "cpu_time": 3.8714925148851603e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0454604577760071e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3458847147177667e+02, + "cpu_time": 7.8367637701501284e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8960816249492671e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7635144277144911e-04, + "cpu_time": 2.0242950156267159e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7633824306982523e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 264, + "real_time": 2.6473341223953124e+06, + "cpu_time": 3.8384524578400588e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0699202214248489e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 264, + "real_time": 2.6468723142787702e+06, + "cpu_time": 3.8377552455676776e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0708047863114304e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 264, + "real_time": 2.6469522968964707e+06, + "cpu_time": 3.8380515782186776e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0706515624542671e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 264, + "real_time": 2.6473571805664422e+06, + "cpu_time": 3.8384211772340095e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0698760630132301e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 264, + "real_time": 2.6467880930850338e+06, + "cpu_time": 3.8374309170079833e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0709661400795776e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6470608014444062e+06, + "cpu_time": 3.8380222751736820e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0704437546566711e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time_median", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6469522968964712e+06, + "cpu_time": 3.8380515782186782e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0706515624542671e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6655929063983461e+02, + "cpu_time": 4.3762082136772813e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1058185038796673e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0070010121958013e-04, + "cpu_time": 1.1402248084866171e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0069766574553772e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 132, + "real_time": 5.2802511773100402e+06, + "cpu_time": 3.8605494419705749e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0837630064551430e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 132, + "real_time": 5.2801723676649006e+06, + "cpu_time": 3.8604744101514548e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0838388845762756e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 132, + "real_time": 5.2806147599987909e+06, + "cpu_time": 3.8609890528782088e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0834129774704765e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 132, + "real_time": 5.2803056879025521e+06, + "cpu_time": 3.8608006353798842e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0837105248470604e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 132, + "real_time": 5.2803347987884823e+06, + "cpu_time": 3.8606171506067926e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0836824979656540e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2803357583329538e+06, + "cpu_time": 3.8606861381973833e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0836815782629227e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time_median", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2803056879025521e+06, + "cpu_time": 3.8606171506067932e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0837105248470604e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6777778436026654e+02, + "cpu_time": 2.0805096297402553e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6152472100084247e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1774074990495568e-05, + "cpu_time": 5.3889633999402990e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1773178259531141e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 66, + "real_time": 1.0546761635465153e+07, + "cpu_time": 3.9213022454545707e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0903863247907936e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 66, + "real_time": 1.0546994423775962e+07, + "cpu_time": 3.9212855272731328e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0902739721729485e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 66, + "real_time": 1.0545813676082727e+07, + "cpu_time": 3.9210501939390087e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0908438977789932e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 66, + "real_time": 1.0546535846184600e+07, + "cpu_time": 3.9209530956053513e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0904953041450356e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 66, + "real_time": 1.0546463005470507e+07, + "cpu_time": 3.9212392450002480e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0905304624073700e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0546513717395790e+07, + "cpu_time": 3.9211660614544630e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0905059922590286e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time_median", + "family_index": 6, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0546535846184600e+07, + "cpu_time": 3.9212392450002480e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0904953041450356e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4318763991663218e+02, + "cpu_time": 1.5568695061178005e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1391982528040735e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.2022193474761519e-05, + "cpu_time": 3.9704248218967724e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2023293088291904e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1078032413215347e+07, + "cpu_time": 4.0782460154552305e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0941273974263046e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1077736534855582e+07, + "cpu_time": 4.0784976678798670e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0941989061509865e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1077843326510806e+07, + "cpu_time": 4.0788096169692588e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0941730962080627e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1078187520756867e+07, + "cpu_time": 4.0787149093938696e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0940899113959709e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1076669972954374e+07, + "cpu_time": 4.0782397027264363e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0944566925317314e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1077693953658596e+07, + "cpu_time": 4.0785015824849325e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0942092007426117e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time_median", + "family_index": 6, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1077843326510806e+07, + "cpu_time": 4.0784976678798664e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0941730962080627e+10, + "dst_id": 6.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.9808551966402342e+02, + "cpu_time": 2.6187040804542608e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4455399921136738e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8375282465860532e-05, + "cpu_time": 6.4207504336892958e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8376141127124290e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57864, + "real_time": 1.2261320953219030e+04, + "cpu_time": 1.1995104359353979e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0878663969137106e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57864, + "real_time": 1.2215047115108433e+04, + "cpu_time": 1.2037790153459558e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0957757885629531e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57864, + "real_time": 1.2056646687197863e+04, + "cpu_time": 1.2043502481439831e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1233101262876775e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57864, + "real_time": 1.2065959232646257e+04, + "cpu_time": 1.2025474479944704e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1216713488253281e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57864, + "real_time": 1.2270656053535607e+04, + "cpu_time": 1.2008381645537813e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0862780187391642e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2173926008341439e+04, + "cpu_time": 1.2022050623947177e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.1029803358657669e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time_median", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2215047115108433e+04, + "cpu_time": 1.2025474479944704e+09, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0957757885629531e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0499617525996925e+02, + "cpu_time": 2.0202598319262874e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8179372884750797e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.6246766398963698e-03, + "cpu_time": 1.6804619237770096e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.6445757835708001e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57260, + "real_time": 1.2062267259013710e+04, + "cpu_time": 1.1909377627438908e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2446414841073953e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57260, + "real_time": 1.2090243100452199e+04, + "cpu_time": 1.1909941371606908e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2348197281562537e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57260, + "real_time": 1.2265118716582387e+04, + "cpu_time": 1.1859909494262547e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1744398226474427e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57260, + "real_time": 1.2226090706455865e+04, + "cpu_time": 1.1918487577314239e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.1877654296286501e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57260, + "real_time": 1.2052405293705440e+04, + "cpu_time": 1.1905075044494975e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2481146918233834e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2139225015241920e+04, + "cpu_time": 1.1900558223023515e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2179562312726259e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time_median", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2090243100452197e+04, + "cpu_time": 1.1909377627438908e+09, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.2348197281562537e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.9063440045838732e+01, + "cpu_time": 2.3237942803517189e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4319000977182813e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.1606066220418026e-03, + "cpu_time": 1.9526767037330826e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1364051913900993e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57542, + "real_time": 1.2363000964643021e+04, + "cpu_time": 1.1963408342829285e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.2827786144200772e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57542, + "real_time": 1.2319398711477887e+04, + "cpu_time": 1.2016497421794815e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.3120939908044979e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57542, + "real_time": 1.2124964716585569e+04, + "cpu_time": 1.2005590709780846e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.4453854005800501e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57542, + "real_time": 1.2174251242707893e+04, + "cpu_time": 1.1977031641747377e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.4111949029584333e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57542, + "real_time": 1.2302809368985216e+04, + "cpu_time": 1.1949645503803053e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.3233021766675040e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2256885000879920e+04, + "cpu_time": 1.1982434723991075e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.3549510170861125e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time_median", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2302809368985214e+04, + "cpu_time": 1.1977031641747375e+09, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 8.3233021766675040e+07, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0186867619841968e+02, + "cpu_time": 2.8119372760778191e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9622162283374253e+05, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.3111391019093806e-03, + "cpu_time": 2.3467161230995856e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.3330425445936121e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56407, + "real_time": 1.2211050594129320e+04, + "cpu_time": 1.1805725370288253e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6771693673799145e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56407, + "real_time": 1.2265140909080741e+04, + "cpu_time": 1.1793962202289803e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6697729077728918e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56407, + "real_time": 1.2437006521411437e+04, + "cpu_time": 1.1757468113880558e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6466985013428929e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56407, + "real_time": 1.2415058185836819e+04, + "cpu_time": 1.1793130112092249e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6496096670222393e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56407, + "real_time": 1.2188009571881243e+04, + "cpu_time": 1.1783925226504412e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6803399996705836e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2303253156467912e+04, + "cpu_time": 1.1786842205011055e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6647180886377046e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time_median", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2265140909080741e+04, + "cpu_time": 1.1793130112092249e+09, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.6697729077728918e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1578564992780589e+02, + "cpu_time": 1.8153181959846478e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5633294383398530e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.4109784180890811e-03, + "cpu_time": 1.5401225912847834e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.3909560364012056e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56390, + "real_time": 1.2608226112879998e+04, + "cpu_time": 1.1819632410792625e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2486727025110298e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56390, + "real_time": 1.2588335664473303e+04, + "cpu_time": 1.1897704231571970e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2538058319811869e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56390, + "real_time": 1.2386701217884238e+04, + "cpu_time": 1.1848288085261846e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3067722615978575e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56390, + "real_time": 1.2423834339336014e+04, + "cpu_time": 1.1814514250356588e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2968887769465452e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56390, + "real_time": 1.2584008613884282e+04, + "cpu_time": 1.1822674104542353e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2549246632593453e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2518221189691567e+04, + "cpu_time": 1.1840562616505075e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2722128472591925e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time_median", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2584008613884283e+04, + "cpu_time": 1.1822674104542353e+09, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.2549246632593453e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0434473089040421e+02, + "cpu_time": 3.4501582538040853e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7363719400037709e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.3354279581135220e-03, + "cpu_time": 2.9138465506653877e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.3624509398762294e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55106, + "real_time": 1.2527314293659114e+04, + "cpu_time": 1.1616005045331209e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.5393106678472197e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55106, + "real_time": 1.2535359619093520e+04, + "cpu_time": 1.1597574214507372e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.5351136695928276e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55106, + "real_time": 1.2744793384641496e+04, + "cpu_time": 1.1611799298411729e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4277228769138145e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55106, + "real_time": 1.2686446467522694e+04, + "cpu_time": 1.1599082713352194e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4572849623190558e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55106, + "real_time": 1.2539349164046393e+04, + "cpu_time": 1.1610995191257651e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.5330344444738925e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2606652585792643e+04, + "cpu_time": 1.1607091292572031e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.4984933242293620e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time_median", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2539349164046393e+04, + "cpu_time": 1.1610995191257648e+09, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.5330344444738925e+08, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0168203644989948e+02, + "cpu_time": 8.2397337342630164e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2217764043666217e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.0657443169721670e-03, + "cpu_time": 7.0988790615750914e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.0353647281554419e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54671, + "real_time": 1.3010120169587006e+04, + "cpu_time": 1.1709714968240168e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2593273379826205e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54671, + "real_time": 1.2968193698725821e+04, + "cpu_time": 1.1754064673092170e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2633987724604850e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54671, + "real_time": 1.2783241923894200e+04, + "cpu_time": 1.1727128253465581e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2816780044955051e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54671, + "real_time": 1.2828029001302923e+04, + "cpu_time": 1.1702453849090815e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2772032241536019e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54671, + "real_time": 1.2978801975023754e+04, + "cpu_time": 1.1684617810003095e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2623661283629386e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2913677353706740e+04, + "cpu_time": 1.1715595910778365e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2687946934910305e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time_median", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2968193698725821e+04, + "cpu_time": 1.1709714968240170e+09, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2633987724604850e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0107361986554180e+02, + "cpu_time": 2.6361821825898848e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9593715167494379e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.8268658180877995e-03, + "cpu_time": 2.2501477540417666e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8494744404602485e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52822, + "real_time": 1.3052703246288356e+04, + "cpu_time": 1.1604156855977399e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.5104378289851837e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52822, + "real_time": 1.3116334298748978e+04, + "cpu_time": 1.1596128788254530e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4982589840764718e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52822, + "real_time": 1.3264898075027966e+04, + "cpu_time": 1.1560201969882269e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4702790639370151e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52822, + "real_time": 1.3264509808740388e+04, + "cpu_time": 1.1598002008872612e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4703513716283865e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52822, + "real_time": 1.3056385716160241e+04, + "cpu_time": 1.1596791209335523e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.5097297760927944e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3150966228993184e+04, + "cpu_time": 1.1591056166464467e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4918114049439702e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time_median", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3116334298748978e+04, + "cpu_time": 1.1596791209335523e+09, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4982589840764718e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0685629112602426e+02, + "cpu_time": 1.7539073861557443e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0209884419564601e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.1253566669834704e-03, + "cpu_time": 1.5131557995812261e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1105192710276691e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51275, + "real_time": 1.3850237830623026e+04, + "cpu_time": 1.1729913866127453e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7317599019923830e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51275, + "real_time": 1.3828300587934074e+04, + "cpu_time": 1.1783289354605396e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7392663750152807e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51275, + "real_time": 1.3632904057401516e+04, + "cpu_time": 1.1753804396622512e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.8071929299920130e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51275, + "real_time": 1.3646947010721093e+04, + "cpu_time": 1.1721640246605091e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.8022462422192068e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51275, + "real_time": 1.3844392028639104e+04, + "cpu_time": 1.1744556044058080e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7337578901572142e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3760556303063762e+04, + "cpu_time": 1.1746640781603708e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7628446678752203e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time_median", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3828300587934074e+04, + "cpu_time": 1.1744556044058080e+09, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7392663750152807e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1052450014564165e+02, + "cpu_time": 2.3997905457003945e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8364995815598361e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.0319790647587087e-03, + "cpu_time": 2.0429589959528528e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.0550592116441170e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 45967, + "real_time": 1.5036246222309126e+04, + "cpu_time": 1.1484713632504022e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.7170692779378548e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 45967, + "real_time": 1.5094347561457811e+04, + "cpu_time": 1.1483944864902146e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.6835154329347553e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 45967, + "real_time": 1.5254153691226189e+04, + "cpu_time": 1.1459963378878610e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.5925448670016594e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 45967, + "real_time": 1.5261293070246367e+04, + "cpu_time": 1.1522716058628798e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.5885251922420540e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 45967, + "real_time": 1.5044764680470191e+04, + "cpu_time": 1.1497586335575452e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.7121336081877251e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5138161045141940e+04, + "cpu_time": 1.1489784854097807e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.6587576756608105e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time_median", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5094347561457811e+04, + "cpu_time": 1.1484713632504022e+09, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.6835154329347553e+09, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1140570334515863e+02, + "cpu_time": 2.2882692648815354e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.3598186758997463e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.3592626616236440e-03, + "cpu_time": 1.9915684183289374e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3449551472918239e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 36242, + "real_time": 1.9416007544839431e+04, + "cpu_time": 1.0337996168812827e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3501436863093676e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 36242, + "real_time": 1.9389902803300942e+04, + "cpu_time": 1.0377152542718867e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3519613927893053e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 36242, + "real_time": 1.9170844812326603e+04, + "cpu_time": 1.0360985234363940e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3674097441519367e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 36242, + "real_time": 1.9304679208668033e+04, + "cpu_time": 1.0367728976370615e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3579298426378109e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 36242, + "real_time": 1.9414077127754696e+04, + "cpu_time": 1.0344308140720508e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3502779363394745e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9339102299377941e+04, + "cpu_time": 1.0357634212597353e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3555445204455790e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time_median", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9389902803300945e+04, + "cpu_time": 1.0360985234363940e+09, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3519613927893053e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0441726867641546e+02, + "cpu_time": 1.6258349919648003e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3521768354306608e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3992820897262699e-03, + "cpu_time": 1.5696972480331436e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4237811628746347e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 31947, + "real_time": 2.1678473731037819e+04, + "cpu_time": 9.4637647950435996e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.4184728431751114e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 31947, + "real_time": 2.1780318048752659e+04, + "cpu_time": 9.4631764662251043e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.4071641140705269e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 31947, + "real_time": 2.1936627645004686e+04, + "cpu_time": 9.4448977888399327e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.3900118490609863e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 31947, + "real_time": 2.1910058816530513e+04, + "cpu_time": 9.4607868167469621e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.3929100528221298e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 31947, + "real_time": 2.1691699018980788e+04, + "cpu_time": 9.4567923703904724e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.4169983159974453e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1799435452061294e+04, + "cpu_time": 9.4578836474492133e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.4051114350252403e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time_median", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1780318048752659e+04, + "cpu_time": 9.4607868167469621e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.4071641140705269e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2006620949068652e+02, + "cpu_time": 7.7597298449354013e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3237351342027180e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.5077669215205924e-03, + "cpu_time": 8.2045097340864375e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5038411731173119e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 21884, + "real_time": 3.2196103528184180e+04, + "cpu_time": 7.6634163706352925e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2568413102600628e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 21884, + "real_time": 3.2121764168556045e+04, + "cpu_time": 7.6704065312368095e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2643786141311932e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 21884, + "real_time": 3.1955071531434205e+04, + "cpu_time": 7.6637489474082005e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2814071436783230e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 21884, + "real_time": 3.1966025098285005e+04, + "cpu_time": 7.6493916606525648e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2802827276021152e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 21884, + "real_time": 3.2222153931578512e+04, + "cpu_time": 7.6578291374768329e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2542082761648327e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2092223651607590e+04, + "cpu_time": 7.6609585294819415e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2674236143673054e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time_median", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2121764168556045e+04, + "cpu_time": 7.6634163706352925e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.2643786141311932e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2577989453786522e+02, + "cpu_time": 7.8529883060271642e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2814037572487947e+08, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.9193262487301822e-03, + "cpu_time": 1.0250660248069779e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9217558189097012e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 13225, + "real_time": 5.2835496152904750e+04, + "cpu_time": 6.0791766719189048e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9692103844939560e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 13225, + "real_time": 5.2617470583790870e+04, + "cpu_time": 6.0840536012497234e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9856571909141533e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 13225, + "real_time": 5.2719944968367607e+04, + "cpu_time": 6.0718039612046552e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9779100703885567e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 13225, + "real_time": 5.2854877314860409e+04, + "cpu_time": 6.0752569955004692e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9677549292321892e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 13225, + "real_time": 5.2819551871365482e+04, + "cpu_time": 6.0776528557537234e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9704085432366325e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2769468178257834e+04, + "cpu_time": 6.0775888171254969e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9741882236530975e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time_median", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2819551871365496e+04, + "cpu_time": 6.0776528557537246e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.9704085432366325e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.9647638464680895e+01, + "cpu_time": 4.5612708657116455e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5140118613114789e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8883578308590552e-03, + "cpu_time": 7.5050665699180673e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8907035697480262e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 7450, + "real_time": 9.3756769931267321e+04, + "cpu_time": 5.0302297388923430e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4736012162906494e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 7450, + "real_time": 9.3772751451063668e+04, + "cpu_time": 5.0249022710917985e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4728387885566551e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 7450, + "real_time": 9.4008696116678941e+04, + "cpu_time": 5.0247299103630519e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4616127797307571e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 7450, + "real_time": 9.3913787163148678e+04, + "cpu_time": 5.0269360597808504e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4661216704141441e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 7450, + "real_time": 9.3730126314987137e+04, + "cpu_time": 5.0312704373385972e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4748728769496429e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.3836426195429158e+04, + "cpu_time": 5.0276136834933281e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4698094663883698e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time_median", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.3772751451063683e+04, + "cpu_time": 5.0269360597808492e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.4728387885566551e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1975068179246666e+02, + "cpu_time": 3.0142442299326404e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7003902926634066e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2761641363351474e-03, + "cpu_time": 5.9953775681473168e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2753094590560597e-03, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3965, + "real_time": 1.7654169817698220e+05, + "cpu_time": 4.3888098663269305e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7516298339842979e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3965, + "real_time": 1.7675341061882506e+05, + "cpu_time": 4.3919234835017216e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7459384068635193e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3965, + "real_time": 1.7667100768114123e+05, + "cpu_time": 4.3912469836059368e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7481520086985069e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3965, + "real_time": 1.7650374526998866e+05, + "cpu_time": 4.3926865075755668e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7526515582818817e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3965, + "real_time": 1.7652729525281326e+05, + "cpu_time": 4.3887166033433354e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7520175211353401e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7659943139995012e+05, + "cpu_time": 4.3906766888706970e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7500778657927094e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time_median", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7654169817698220e+05, + "cpu_time": 4.3912469836059368e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7516298339842979e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0784867649877492e+02, + "cpu_time": 1.8197572562559193e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8999485733812593e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.1069662367443727e-04, + "cpu_time": 4.1445940687652162e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1050548123958086e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2051, + "real_time": 3.4132509869850287e+05, + "cpu_time": 4.0890029673425305e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9153185816023293e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2051, + "real_time": 3.4134241861362976e+05, + "cpu_time": 4.0908026195324856e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9150691754458931e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2051, + "real_time": 3.4109098584026098e+05, + "cpu_time": 4.0903192836968690e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9186922834299332e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2051, + "real_time": 3.4112554959755763e+05, + "cpu_time": 4.0881436077077699e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9181939083111473e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2051, + "real_time": 3.4134343873765290e+05, + "cpu_time": 4.0890347023150301e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9150544864858238e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4124549829752080e+05, + "cpu_time": 4.0894606361189371e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9164656870550262e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time_median", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4132509869850281e+05, + "cpu_time": 4.0890347023150301e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.9153185816023293e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2607923230310855e+02, + "cpu_time": 1.0797730451606397e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8167321421057094e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6946782575043429e-04, + "cpu_time": 2.6403800922397144e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6951994740635238e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1042, + "real_time": 6.7108008978600590e+05, + "cpu_time": 3.9286433894527549e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0000637048701355e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1042, + "real_time": 6.7102588739312266e+05, + "cpu_time": 3.9293081252108610e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0004675870786530e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1042, + "real_time": 6.7089904463249340e+05, + "cpu_time": 3.9300579630131418e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0014129947644394e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1042, + "real_time": 6.7088015982166154e+05, + "cpu_time": 3.9287774399618590e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0015537810686920e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1042, + "real_time": 6.7108638805601804e+05, + "cpu_time": 3.9289668673503494e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0000167783464394e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7099431393786031e+05, + "cpu_time": 3.9291507569977933e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0007029692256714e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time_median", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7102588739312266e+05, + "cpu_time": 3.9289668673503494e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0004675870786530e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8665375120839698e+01, + "cpu_time": 5.6532962311700416e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3534662950420873e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4704353385918103e-04, + "cpu_time": 1.4388086843197744e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4704865176546822e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 526, + "real_time": 1.3303830042240624e+06, + "cpu_time": 3.8614901638209045e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0443266177427475e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 526, + "real_time": 1.3307693202806346e+06, + "cpu_time": 3.8626120356854892e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0428622735191986e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 526, + "real_time": 1.3300973697151181e+06, + "cpu_time": 3.8614687126810324e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0454098720887970e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 526, + "real_time": 1.3301840617264772e+06, + "cpu_time": 3.8612360276034796e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0450810478737671e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 526, + "real_time": 1.3308550971259181e+06, + "cpu_time": 3.8625856807595420e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0425372487903938e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3304577706144422e+06, + "cpu_time": 3.8618785241100895e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0440434120029816e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time_median", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3303830042240624e+06, + "cpu_time": 3.8614901638209045e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0443266177427475e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4107437828729292e+02, + "cpu_time": 6.6514671816431117e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2930315705472933e+07, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5635866528087942e-04, + "cpu_time": 1.7223398250663100e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5634822402010861e-04, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 264, + "real_time": 2.6489879061101060e+06, + "cpu_time": 3.8357298629920614e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0667550308710693e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 264, + "real_time": 2.6486818564025629e+06, + "cpu_time": 3.8355902116658980e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0673404839301613e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 264, + "real_time": 2.6485042824678011e+06, + "cpu_time": 3.8358185743935239e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0676802332726357e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 264, + "real_time": 2.6488363977422882e+06, + "cpu_time": 3.8359888695462167e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0670448395529175e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 264, + "real_time": 2.6488224667087761e+06, + "cpu_time": 3.8357096026887250e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0670714888177711e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6487665818863073e+06, + "cpu_time": 3.8357674242572850e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0671784152889107e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time_median", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6488224667087761e+06, + "cpu_time": 3.8357298629920614e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0670714888177711e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8231920241527712e+02, + "cpu_time": 1.4816615531140738e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4878991486337446e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.8831736122795429e-05, + "cpu_time": 3.8627512808625663e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8833162418554351e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 132, + "real_time": 5.2852444527106304e+06, + "cpu_time": 3.8619102686366236e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0789600822025963e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 132, + "real_time": 5.2855754320537951e+06, + "cpu_time": 3.8621979740918797e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0786420409801071e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 132, + "real_time": 5.2857744705044860e+06, + "cpu_time": 3.8622229479549193e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0784508022034454e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 132, + "real_time": 5.2853196252588974e+06, + "cpu_time": 3.8617221663626868e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0788878446845284e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 132, + "real_time": 5.2856846155147208e+06, + "cpu_time": 3.8621104850011539e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0785371342830246e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2855197192085069e+06, + "cpu_time": 3.8620327684094530e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0786955808707397e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time_median", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2855754320537951e+06, + "cpu_time": 3.8621104850011533e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0786420409801071e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2967485611413341e+02, + "cpu_time": 2.1270923355719155e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2068915014562905e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.3453599327130430e-05, + "cpu_time": 5.5077014182040235e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3453903986068800e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 66, + "real_time": 1.0557830291376872e+07, + "cpu_time": 3.9239518540909320e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0850496473550095e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 66, + "real_time": 1.0558501676176533e+07, + "cpu_time": 3.9244022716674423e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0847263036511902e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 66, + "real_time": 1.0558858062281754e+07, + "cpu_time": 3.9244472003025001e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0845546822700920e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 66, + "real_time": 1.0559415868060155e+07, + "cpu_time": 3.9244804722725898e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0842860884370796e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 66, + "real_time": 1.0558499376091992e+07, + "cpu_time": 3.9244892475764549e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0847274113180984e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0558621054797463e+07, + "cpu_time": 3.9243542091819847e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0846688266062943e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time_median", + "family_index": 7, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0558501676176533e+07, + "cpu_time": 3.9244472003025007e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.0847263036511902e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.7928067800168458e+02, + "cpu_time": 2.2749984076921090e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7896131000552746e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.4863288965037723e-05, + "cpu_time": 5.7971281042093370e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4863221090392447e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1103297332019515e+07, + "cpu_time": 4.0820724775767994e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0880286957376930e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1099461919882081e+07, + "cpu_time": 4.0810997533343184e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0889535860069023e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1103117671428304e+07, + "cpu_time": 4.0822006721210724e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0880720124768501e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1102395077997986e+07, + "cpu_time": 4.0821474166666645e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0882462394968452e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1101764995943416e+07, + "cpu_time": 4.0820526775759751e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0883981705151924e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1102007399454266e+07, + "cpu_time": 4.0819145994549662e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0883397408466965e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time_median", + "family_index": 7, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1102395077997990e+07, + "cpu_time": 4.0820724775767994e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.0882462394968452e+10, + "dst_id": 7.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5479946901680989e+03, + "cpu_time": 4.5934917887431780e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7329277975551574e+06, + "dst_id": 0.0000000000000000e+00, + "src_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.3357698197382528e-05, + "cpu_time": 1.1253277541270754e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3362392994104605e-05, + "dst_id": 0.0000000000000000e+00, + "src_id": NaN + } + ] +} diff --git a/results/hipMemcpyAsync_GPUToPageable.json b/results/hipMemcpyAsync_GPUToPageable.json new file mode 100644 index 0000000..7ce55b5 --- /dev/null +++ b/results/hipMemcpyAsync_GPUToPageable.json @@ -0,0 +1,3858 @@ +{ + "context": { + "date": "2023-10-09T17:44:05-04:00", + "host_name": "frontier08191", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1813, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [2.29,5.25,19.34], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 92456, + "real_time": 7.5790847097529013e+03, + "cpu_time": 1.4576654646534562e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.3777165687378399e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 92456, + "real_time": 7.5816833849294026e+03, + "cpu_time": 1.4619616628450292e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.3765588326844089e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 92456, + "real_time": 7.6116243127170292e+03, + "cpu_time": 1.4589161341611138e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.3632768707763352e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 92456, + "real_time": 7.5651537838648774e+03, + "cpu_time": 1.4600630483689552e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.3839364977087744e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 92456, + "real_time": 7.6124390794376586e+03, + "cpu_time": 1.4578567686250759e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.3629168959984779e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5899970541403745e+03, + "cpu_time": 1.4592926157307264e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.3728811331811674e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5816833849294026e+03, + "cpu_time": 1.4589161341611136e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.3765588326844089e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1075858504034521e+01, + "cpu_time": 1.7725966022285000e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.3629038559426102e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7767940295230489e-03, + "cpu_time": 1.2146957937842302e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7759365024263073e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 86703, + "real_time": 8.0256454274149437e+03, + "cpu_time": 1.4984271051751386e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 6.3795492167028725e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 86703, + "real_time": 8.0624378568289985e+03, + "cpu_time": 1.5054889092649621e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 6.3504365440327555e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 86703, + "real_time": 8.0416331759580216e+03, + "cpu_time": 1.4995567754287629e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 6.3668658939917892e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 86703, + "real_time": 8.0506317142925518e+03, + "cpu_time": 1.5061534202968744e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 6.3597493733445734e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 86703, + "real_time": 8.0281549701267159e+03, + "cpu_time": 1.4982953750158593e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 6.3775550161299221e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0417006289242472e+03, + "cpu_time": 1.5015843170363198e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 6.3668312088403821e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0416331759580207e+03, + "cpu_time": 1.4995567754287627e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 6.3668658939917892e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5419596497807794e+01, + "cpu_time": 3.9057200487121108e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2203343783570141e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9174546789701250e-03, + "cpu_time": 2.6010660902618106e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9167060321350637e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 77804, + "real_time": 9.0120509103686200e+03, + "cpu_time": 1.6001098465374545e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1362563418520626e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 77804, + "real_time": 9.0628142519478624e+03, + "cpu_time": 1.6097592771579843e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1298918542658123e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 77804, + "real_time": 8.8521144319233172e+03, + "cpu_time": 1.5873780152691381e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1567857689538626e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 77804, + "real_time": 9.0052658891548090e+03, + "cpu_time": 1.6071959629324974e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1371124546508063e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 77804, + "real_time": 8.8639465751894440e+03, + "cpu_time": 1.5857706994499016e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1552416198741750e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9592384117168094e+03, + "cpu_time": 1.5980427602693951e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1430576079193437e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0052658891548090e+03, + "cpu_time": 1.6001098465374545e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1371124546508063e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5120613602257180e+01, + "cpu_time": 1.1064255457401752e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2164001083905329e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0617042345681895e-02, + "cpu_time": 6.9236291621737073e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0641634332014913e-02, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 72804, + "real_time": 9.6116098268893584e+03, + "cpu_time": 1.6555067784737090e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.1307564881280681e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 72804, + "real_time": 9.6171790300429329e+03, + "cpu_time": 1.6589679454425572e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.1295225903586587e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 72804, + "real_time": 9.6082037916300433e+03, + "cpu_time": 1.6573272924564611e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.1315118251176834e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 72804, + "real_time": 9.6134300429252271e+03, + "cpu_time": 1.6591714067908386e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.1303530486573589e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 72804, + "real_time": 9.6068437861344755e+03, + "cpu_time": 1.6533772965771117e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.1318135753970221e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6114532955244085e+03, + "cpu_time": 1.6568701439481352e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.1307915055317581e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6116098268893584e+03, + "cpu_time": 1.6573272924564611e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.1307564881280681e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1389286176922591e+00, + "cpu_time": 2.4476680081007952e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.1748344453728423e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.3062464025284911e-04, + "cpu_time": 1.4772841535235085e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3058339689988487e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 68983, + "real_time": 1.0153305496328829e+04, + "cpu_time": 1.7211286244437077e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.0341541988281620e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 68983, + "real_time": 1.0149266814589211e+04, + "cpu_time": 1.7227843816592529e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.0357595034472293e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 68983, + "real_time": 1.0151713988343079e+04, + "cpu_time": 1.7249032399286760e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.0347866426332724e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 68983, + "real_time": 1.0150142049411486e+04, + "cpu_time": 1.7232608729687014e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.0354115046473551e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 68983, + "real_time": 1.0151245232640560e+04, + "cpu_time": 1.7195475305510125e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.0349729576324511e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0151134716262633e+04, + "cpu_time": 1.7223249299102699e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.0350169614376938e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0151245232640558e+04, + "cpu_time": 1.7227843816592525e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.0349729576324511e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5438993049526197e+00, + "cpu_time": 2.0542554515035135e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1367243053603117e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5209130290421764e-04, + "cpu_time": 1.1927223579179897e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5208670407109690e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 62388, + "real_time": 1.1223158434911607e+04, + "cpu_time": 1.8536376146053793e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.2991930457983577e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 62388, + "real_time": 1.1222678391105888e+04, + "cpu_time": 1.8604921250881551e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.2995052647078109e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 62388, + "real_time": 1.1222977772218122e+04, + "cpu_time": 1.8539498300955343e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.2993105450844383e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 62388, + "real_time": 1.1221621553335050e+04, + "cpu_time": 1.8537336138359940e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.3001927226509881e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 62388, + "real_time": 1.1220793875938703e+04, + "cpu_time": 1.8476496537795760e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.3007312054510748e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1222246005501875e+04, + "cpu_time": 1.8538925674809278e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.2997865567385340e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1222678391105886e+04, + "cpu_time": 1.8537336138359940e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 7.2995052647078109e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0066762111097884e+00, + "cpu_time": 4.5448233829724771e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5484721882283353e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.9703630682864215e-05, + "cpu_time": 2.4515031036280543e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9707721415270011e-05, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 62366, + "real_time": 1.1229213263756144e+04, + "cpu_time": 1.8930920212936504e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.4590514593646243e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 62366, + "real_time": 1.1224700536383602e+04, + "cpu_time": 1.9027134127569490e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.4596380497540321e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 62366, + "real_time": 1.1229842887239221e+04, + "cpu_time": 1.8938252252830123e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.4589696547417943e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 62366, + "real_time": 1.1228299274867808e+04, + "cpu_time": 1.9015867780521476e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.4591702268456759e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 62366, + "real_time": 1.1225591083172145e+04, + "cpu_time": 1.8887089327518184e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.4595222539827440e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1227529409083783e+04, + "cpu_time": 1.8959852740275157e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.4592703289377744e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1228299274867808e+04, + "cpu_time": 1.8938252252830127e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.4591702268456759e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2660466189057691e+00, + "cpu_time": 5.9713181631452372e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9453951911415893e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0182949750925555e-04, + "cpu_time": 3.1494538723187246e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0184027131460889e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 60920, + "real_time": 1.1497334656440022e+04, + "cpu_time": 1.9541514757058489e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.8500518580317745e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 60920, + "real_time": 1.1488940743384974e+04, + "cpu_time": 1.9560752757715069e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.8521341289767675e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 60920, + "real_time": 1.1498668239996587e+04, + "cpu_time": 1.9526032337491772e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.8497213169452858e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 60920, + "real_time": 1.1490760985657782e+04, + "cpu_time": 1.9554910045961860e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.8516823246867156e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 60920, + "real_time": 1.1500000988525939e+04, + "cpu_time": 1.9539700952068200e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.8493910594176545e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1495141122801058e+04, + "cpu_time": 1.9544582170059079e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.8505961376116395e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1497334656440024e+04, + "cpu_time": 1.9541514757058489e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.8500518580317745e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9623771191960389e+00, + "cpu_time": 1.3651781257084803e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2307502507982682e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.3169344910024384e-04, + "cpu_time": 6.9849440311895599e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3175188324974279e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56362, + "real_time": 1.2422378913991202e+04, + "cpu_time": 2.2860644263865794e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2756400729483023e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56362, + "real_time": 1.2420636379492313e+04, + "cpu_time": 2.2947397803484600e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2763802109372072e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56362, + "real_time": 1.2421805903266953e+04, + "cpu_time": 2.2860791277811168e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2758834351745863e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56362, + "real_time": 1.2409391844215852e+04, + "cpu_time": 2.2932426315602628e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2811613028842354e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56362, + "real_time": 1.2417686964678391e+04, + "cpu_time": 2.2897056740356937e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2776334422356195e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2418380001128944e+04, + "cpu_time": 2.2899663280224224e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2773396928359909e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2420636379492311e+04, + "cpu_time": 2.2897056740356933e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2763802109372072e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3405668451314554e+00, + "cpu_time": 3.9975966082749053e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2705171048023398e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.3005342441171471e-04, + "cpu_time": 1.7457010434416146e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3023895313856254e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 48140, + "real_time": 1.4540780548305140e+04, + "cpu_time": 2.8081541773992525e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.0140965654885445e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 48140, + "real_time": 1.4537620367325238e+04, + "cpu_time": 2.8189282322393115e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.0160560454995422e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 48140, + "real_time": 1.4541409611934634e+04, + "cpu_time": 2.8086638803489644e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.0137066142765636e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 48140, + "real_time": 1.4541365656040582e+04, + "cpu_time": 2.8207557582052163e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.0137338610663300e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 48140, + "real_time": 1.4544101428761627e+04, + "cpu_time": 2.8107490112172756e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.0120383608436012e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4541055522473445e+04, + "cpu_time": 2.8134502118820041e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.0139262894349155e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4541365656040583e+04, + "cpu_time": 2.8107490112172753e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.0137338610663300e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3116557415858900e+00, + "cpu_time": 5.9504863231144768e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4330400477865229e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5897441131513374e-04, + "cpu_time": 2.1150139064071132e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5898067077231006e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 35855, + "real_time": 1.9529382761835303e+04, + "cpu_time": 4.1785340036256901e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3423056078980993e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 35855, + "real_time": 1.9520891090427915e+04, + "cpu_time": 4.1845148235950619e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3428895165986687e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 35855, + "real_time": 1.9529044016037522e+04, + "cpu_time": 4.1858196569515967e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3423288911875240e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 35855, + "real_time": 1.9519414003508180e+04, + "cpu_time": 4.1849271705480503e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3429911366851763e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 35855, + "real_time": 1.9528293825591802e+04, + "cpu_time": 4.1836569320875940e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3423804575106333e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9525405139480146e+04, + "cpu_time": 4.1834905173615989e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3425791219760204e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9528293825591805e+04, + "cpu_time": 4.1845148235950626e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3423804575106333e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8393635660650345e+00, + "cpu_time": 2.8780544268841549e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3278893827770175e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4784958527082732e-04, + "cpu_time": 6.8795528875711595e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4787286859332352e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 20705, + "real_time": 3.3825512705809910e+04, + "cpu_time": 7.0190524559285506e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5499779842507687e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 20705, + "real_time": 3.3808667346366157e+04, + "cpu_time": 7.0174407341221755e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5507502695351044e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 20705, + "real_time": 3.3823488700267466e+04, + "cpu_time": 7.0103131417532219e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5500707352989700e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 20705, + "real_time": 3.3809648615288599e+04, + "cpu_time": 7.0248737551316182e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5507052615829878e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 20705, + "real_time": 3.3823749215393946e+04, + "cpu_time": 7.0156349335908570e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5500587964428993e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3818213316625217e+04, + "cpu_time": 7.0174630041052849e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5503126094221460e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3823488700267466e+04, + "cpu_time": 7.0174407341221755e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5500707352989700e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.3101833119278545e+00, + "cpu_time": 5.2893396869179718e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8099219727863194e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4573099809038475e-04, + "cpu_time": 7.5373958990929580e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4575185350562338e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 12172, + "real_time": 5.7389009811667289e+04, + "cpu_time": 1.2138089558002025e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.8271372923859417e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 12172, + "real_time": 5.7378046975335339e+04, + "cpu_time": 1.2133329485704898e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.8274863911815319e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 12172, + "real_time": 5.7364995689901436e+04, + "cpu_time": 1.2131230997370994e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.8279021681938206e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 12172, + "real_time": 5.7362470281347734e+04, + "cpu_time": 1.2127877842589546e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.8279826424088993e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 12172, + "real_time": 5.7331626985635929e+04, + "cpu_time": 1.2128011666118968e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.8289660613725719e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.7365229948777545e+04, + "cpu_time": 1.2131707909957285e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.8278949111085533e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.7364995689901421e+04, + "cpu_time": 1.2131230997370994e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.8279021681938206e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1602012760664660e+01, + "cpu_time": 4.2401067167406730e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8847421647582417e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7656979288592569e-04, + "cpu_time": 3.4950616584335502e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7664868603320804e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5772, + "real_time": 1.2142500338832644e+05, + "cpu_time": 2.2952624255024263e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.7271171023097668e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5772, + "real_time": 1.2159367731390927e+05, + "cpu_time": 2.2984171309771284e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.7247212571636765e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5772, + "real_time": 1.2143694759631211e+05, + "cpu_time": 2.2945291545391487e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.7269472277675137e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5772, + "real_time": 1.2140315721087881e+05, + "cpu_time": 2.3011653915453877e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.7274278924700619e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5772, + "real_time": 1.2158322873119844e+05, + "cpu_time": 2.3011627702702896e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.7248694757370491e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2148840284812501e+05, + "cpu_time": 2.2981073745668764e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.7262165910896137e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2143694759631211e+05, + "cpu_time": 2.2984171309771287e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.7269472277675137e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.2207123306846228e+01, + "cpu_time": 3.1496132112464710e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3098420487762637e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5897880904826881e-04, + "cpu_time": 1.3705248266913889e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5879356943816178e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3041, + "real_time": 2.3011765948205662e+05, + "cpu_time": 4.3770410194014863e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.8226780202095051e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3041, + "real_time": 2.2997456708773988e+05, + "cpu_time": 4.3721050575468806e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.8238121080579269e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3041, + "real_time": 2.3002311332490930e+05, + "cpu_time": 4.3726530450509442e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.8234271936296745e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3041, + "real_time": 2.3004430105225375e+05, + "cpu_time": 4.3784767609338922e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.8232592508550247e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3041, + "real_time": 2.3005333022614595e+05, + "cpu_time": 4.3727290562314924e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.8231876912526913e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3004259423462115e+05, + "cpu_time": 4.3746009878329403e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.8232728528009644e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3004430105225378e+05, + "cpu_time": 4.3727290562314930e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.8232592508550247e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1865606671540171e+01, + "cpu_time": 2.9369903036530445e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1105980265260674e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2546088407715607e-04, + "cpu_time": 6.7137330051853514e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2545161138176593e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1547, + "real_time": 4.5247165230433951e+05, + "cpu_time": 8.8998479508726951e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.8539521663464767e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1547, + "real_time": 4.5232504983863031e+05, + "cpu_time": 8.8989821331608703e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.8545530482984936e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1547, + "real_time": 4.5265044065288064e+05, + "cpu_time": 8.8999207498383510e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.8532198903641155e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1547, + "real_time": 4.5242682843587681e+05, + "cpu_time": 8.9007357207498897e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.8541358453478474e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1547, + "real_time": 4.5235133582756825e+05, + "cpu_time": 8.8986502327084553e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.8544452808242954e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5244506141185900e+05, + "cpu_time": 8.8996273574660509e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.8540612462362457e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5242682843587676e+05, + "cpu_time": 8.8998479508726951e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.8541358453478474e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2884817858636015e+02, + "cpu_time": 8.2678547261268093e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2789558356629582e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8478193172070039e-04, + "cpu_time": 9.2901133879395089e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8472391871515930e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 786, + "real_time": 8.8906707960470044e+05, + "cpu_time": 1.7900090470737834e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8870585116546585e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 786, + "real_time": 8.8857141078708100e+05, + "cpu_time": 1.7867276997455508e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8881111631916042e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 786, + "real_time": 8.8876398646393581e+05, + "cpu_time": 1.7862251437659049e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8877020508842125e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 786, + "real_time": 8.8850160387970274e+05, + "cpu_time": 1.7871635826972001e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8882595064253281e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 786, + "real_time": 8.8880553540332918e+05, + "cpu_time": 1.7869891972010205e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8876138065889412e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8874192322774976e+05, + "cpu_time": 1.7874229340966921e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8877490077489491e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8876398646393570e+05, + "cpu_time": 1.7869891972010203e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8877020508842125e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2197288429959568e+02, + "cpu_time": 1.4884071729154580e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7144316138088163e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4976078937902500e-04, + "cpu_time": 8.3271124283053506e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4973826469815245e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 71, + "real_time": 9.6594816677167378e+06, + "cpu_time": 1.8739816633802895e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4737300772714686e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 71, + "real_time": 9.6169690614644904e+06, + "cpu_time": 1.8676452802817035e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4890859880639219e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 71, + "real_time": 9.6358986125445701e+06, + "cpu_time": 1.8689815225352217e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4822317408276691e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 71, + "real_time": 9.6315473029521145e+06, + "cpu_time": 1.8714126225352030e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4838049323305931e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 71, + "real_time": 9.6640340118131172e+06, + "cpu_time": 1.8708798591549240e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4720937404590826e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6415861312982086e+06, + "cpu_time": 1.8705801895774685e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4801892957905474e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6358986125445701e+06, + "cpu_time": 1.8708798591549240e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4822317408276691e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9769190895899803e+04, + "cpu_time": 2.4249845052857865e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1353882641829662e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0504085766267950e-03, + "cpu_time": 1.2963809404148284e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0502874004047866e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 37, + "real_time": 1.8686962842538550e+07, + "cpu_time": 3.6070114945946105e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5912130058521338e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 37, + "real_time": 1.8743001931422465e+07, + "cpu_time": 3.6117252972972564e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5804757554601016e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 37, + "real_time": 1.8683265944993176e+07, + "cpu_time": 3.6069822783784159e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5919236068030248e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 37, + "real_time": 1.8715438558845907e+07, + "cpu_time": 3.6105855486486435e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5857489413883276e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 37, + "real_time": 1.8644848586739719e+07, + "cpu_time": 3.6049190567566894e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5993246975321674e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8694703572907966e+07, + "cpu_time": 3.6082447351351239e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5897372014071503e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8686962842538550e+07, + "cpu_time": 3.6070114945946112e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5912130058521338e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6874547326261803e+04, + "cpu_time": 2.8181792362840108e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.0809435865681069e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9724595890196271e-03, + "cpu_time": 7.8103882722868384e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9725520809134524e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 19, + "real_time": 3.6870041960164122e+07, + "cpu_time": 7.0217140631578892e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6402922498708911e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 19, + "real_time": 3.6919784585112020e+07, + "cpu_time": 7.0309112263158262e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6353876250438247e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 19, + "real_time": 3.6856584839130700e+07, + "cpu_time": 7.0264863631579325e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6416213977997441e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 19, + "real_time": 3.6834471515919030e+07, + "cpu_time": 7.0236868894737288e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6438076203155003e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 19, + "real_time": 3.6890967896110132e+07, + "cpu_time": 7.0260313421053857e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6382273400355058e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6874370159287199e+07, + "cpu_time": 7.0257659768421531e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6398672466130939e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6870041960164115e+07, + "cpu_time": 7.0260313421053857e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6402922498708911e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2653130035717513e+04, + "cpu_time": 3.4570065094420679e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2225968579392615e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.8552373626084779e-04, + "cpu_time": 4.9204691998521084e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8536109687458971e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10, + "real_time": 7.3235759884119034e+07, + "cpu_time": 1.3668543430000001e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6653604253543010e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10, + "real_time": 7.3415408283472061e+07, + "cpu_time": 1.3687273029999858e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6563912436952639e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10, + "real_time": 7.3812286555767059e+07, + "cpu_time": 1.3725559290000147e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6367313427851906e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10, + "real_time": 7.3454286903142929e+07, + "cpu_time": 1.3682426740000153e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6544559523661280e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10, + "real_time": 7.3310846835374832e+07, + "cpu_time": 1.3669007759999943e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6616062641152210e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.3445717692375198e+07, + "cpu_time": 1.3686562050000021e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6549090456632204e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.3415408283472061e+07, + "cpu_time": 1.3682426740000153e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6563912436952639e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2228205762252235e+05, + "cpu_time": 2.3298147850056883e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1033077074499208e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.0264808433562180e-03, + "cpu_time": 1.7022644375514922e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0187008586686033e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5, + "real_time": 1.4578309953212738e+08, + "cpu_time": 2.6130130959999746e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6826690729105091e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5, + "real_time": 1.4578041434288025e+08, + "cpu_time": 2.6134430019999966e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6827369055026984e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5, + "real_time": 1.4533545374870300e+08, + "cpu_time": 2.6072637179999560e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6940120125698586e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5, + "real_time": 1.4592687487602234e+08, + "cpu_time": 2.6141203420000353e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6790407007353435e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5, + "real_time": 1.4513334035873413e+08, + "cpu_time": 2.6049142300000200e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6991563115200572e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4559183657169345e+08, + "cpu_time": 2.6105508775999966e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6875230006476936e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4578041434288025e+08, + "cpu_time": 2.6130130959999746e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6827369055026984e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3924352992705296e+05, + "cpu_time": 4.1756793907135032e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.6002606150651034e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3300999418329345e-03, + "cpu_time": 1.5995395556329488e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3322595177181307e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 2.8426723182201385e+08, + "cpu_time": 4.5065636299999315e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7772268619138417e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 2.8468500077724457e+08, + "cpu_time": 4.5155940200000089e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7716838648628454e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 2.8438818454742432e+08, + "cpu_time": 4.5085951300001168e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7756203750474162e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 2.8548514842987061e+08, + "cpu_time": 4.5197412599999607e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7611127230450816e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 2.8418129682540894e+08, + "cpu_time": 4.5090414599999917e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7783690763423796e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8460137248039246e+08, + "cpu_time": 4.5119071000000018e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7728025802423134e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8438818454742432e+08, + "cpu_time": 4.5090414599999917e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7756203750474162e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2953192802598490e+05, + "cpu_time": 5.5385275909565133e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.0072177090311674e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPageable/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8606091861432146e-03, + "cpu_time": 1.2275358220377609e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8572977408696321e-03, + "hip_id": NaN, + "numa_id": NaN + } + ] +} diff --git a/results/hipMemcpyAsync_GPUToPinned.json b/results/hipMemcpyAsync_GPUToPinned.json new file mode 100644 index 0000000..a1895e2 --- /dev/null +++ b/results/hipMemcpyAsync_GPUToPinned.json @@ -0,0 +1,3858 @@ +{ + "context": { + "date": "2023-10-09T17:38:29-04:00", + "host_name": "frontier08191", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1870, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.31,11.54,26.8], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53295, + "real_time": 1.3094350788613152e+04, + "cpu_time": 1.9975676836476214e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9550415605378281e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53295, + "real_time": 1.3044469961807499e+04, + "cpu_time": 1.9867436964067925e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9625174556692183e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53295, + "real_time": 1.3008615265856624e+04, + "cpu_time": 1.9804497982925233e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9679265991662968e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53295, + "real_time": 1.3032896488822427e+04, + "cpu_time": 1.9848681264658979e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9642602104571044e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53295, + "real_time": 1.3044607950029651e+04, + "cpu_time": 1.9856583713293941e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9624966958046302e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3044988091025873e+04, + "cpu_time": 1.9870575352284461e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9624485043270156e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3044469961807499e+04, + "cpu_time": 1.9856583713293941e+04, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9625174556692183e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1250112529783308e+01, + "cpu_time": 6.3437976078423318e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6947229729738843e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3955646652741224e-03, + "cpu_time": 3.1925585924783024e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3922783006139823e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53884, + "real_time": 1.2985968137719734e+04, + "cpu_time": 1.9807208800386023e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9427172049869552e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53884, + "real_time": 1.3013354082871958e+04, + "cpu_time": 1.9893156354390892e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9344199561425067e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53884, + "real_time": 1.3059652373227924e+04, + "cpu_time": 1.9917366064137779e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9204718882838853e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53884, + "real_time": 1.3009576574422315e+04, + "cpu_time": 1.9863520915299534e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9355623687755197e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53884, + "real_time": 1.3072530672963361e+04, + "cpu_time": 1.9899500185583827e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9166096665500246e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3028216368241059e+04, + "cpu_time": 1.9876150463959610e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9299562169477783e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3013354082871960e+04, + "cpu_time": 1.9893156354390892e+04, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9344199561425067e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6418478944675762e+01, + "cpu_time": 4.3143927680100916e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0980432816919149e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7953541693898523e-03, + "cpu_time": 2.1706380095246089e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7940343888734620e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53587, + "real_time": 1.3115991310715624e+04, + "cpu_time": 1.9936710974676691e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.8072634827335015e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53587, + "real_time": 1.3093081719545800e+04, + "cpu_time": 1.9929239554369535e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.8209242249770552e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53587, + "real_time": 1.3169547687583350e+04, + "cpu_time": 2.0019459383805839e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.7755138163587674e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53587, + "real_time": 1.3077859703108723e+04, + "cpu_time": 1.9926110175975489e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.8300274146279916e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53587, + "real_time": 1.3064396643654294e+04, + "cpu_time": 1.9903323828540510e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.8380963769756839e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3104175412921559e+04, + "cpu_time": 1.9942968783473618e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.8143650631346017e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3093081719545800e+04, + "cpu_time": 1.9929239554369535e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.8209242249770552e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1264654472746514e+01, + "cpu_time": 4.4537621568198752e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4552794826149629e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1489699406844715e-03, + "cpu_time": 2.2332493246996548e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1420076522891147e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52954, + "real_time": 1.3206203769456990e+04, + "cpu_time": 2.0105202911961285e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5507863090349767e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52954, + "real_time": 1.3212322023701336e+04, + "cpu_time": 2.0139520602787336e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5500681835684383e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52954, + "real_time": 1.3278802913989846e+04, + "cpu_time": 2.0244325546700864e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5423077014286697e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52954, + "real_time": 1.3234845188806737e+04, + "cpu_time": 2.0189492163009425e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5474302651700673e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52954, + "real_time": 1.3207116621340483e+04, + "cpu_time": 2.0173209011594961e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5506791215054283e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3227858103459081e+04, + "cpu_time": 2.0170350047210773e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5482543161415163e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3212322023701336e+04, + "cpu_time": 2.0173209011594961e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5500681835684383e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0759885066734210e+01, + "cpu_time": 5.2539443929710941e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5924250392318558e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3253866821182872e-03, + "cpu_time": 2.6047859262103524e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3203068137957607e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52173, + "real_time": 1.3469463507500583e+04, + "cpu_time": 2.0487335192532581e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0409525945254678e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52173, + "real_time": 1.3472897656609119e+04, + "cpu_time": 2.0464076878845386e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0401774765881270e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52173, + "real_time": 1.3386572877022310e+04, + "cpu_time": 2.0454900331589128e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0597823936182153e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52173, + "real_time": 1.3486499356167817e+04, + "cpu_time": 2.0540076878845364e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0371113302480274e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52173, + "real_time": 1.3432674862050941e+04, + "cpu_time": 2.0477974277883230e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0492809824287003e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3449621651870155e+04, + "cpu_time": 2.0484872711939141e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0454609554817075e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3469463507500583e+04, + "cpu_time": 2.0477974277883230e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0409525945254678e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0490055571213432e+01, + "cpu_time": 3.3286493149299176e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.1876670794059581e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.0104977388403587e-03, + "cpu_time": 1.6249304361017048e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0168395568718510e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51470, + "real_time": 1.3684937415842163e+04, + "cpu_time": 2.1006997998834260e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 5.9861435613995969e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51470, + "real_time": 1.3606532480538341e+04, + "cpu_time": 2.0973855197202283e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.0206375222468758e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51470, + "real_time": 1.3672969838944475e+04, + "cpu_time": 2.1114219545366272e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 5.9913830692925787e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51470, + "real_time": 1.3619529295040460e+04, + "cpu_time": 2.0995680493491247e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.0148921614957058e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51470, + "real_time": 1.3615037491795943e+04, + "cpu_time": 2.0980609850398374e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.0168765638260484e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3639801304432276e+04, + "cpu_time": 2.1014272617058487e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.0059865756521606e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3619529295040460e+04, + "cpu_time": 2.0995680493491247e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.0148921614957058e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6291995189144345e+01, + "cpu_time": 5.7347025411590053e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5965268165865552e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6607422189758143e-03, + "cpu_time": 2.7289560032184123e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6582257493860551e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50185, + "real_time": 1.3922363749492280e+04, + "cpu_time": 2.1785969193982266e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.1768116603473666e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50185, + "real_time": 1.3932948791715640e+04, + "cpu_time": 2.1864787187406568e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.1759176212390680e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50185, + "real_time": 1.3950302043939966e+04, + "cpu_time": 2.1885805200757193e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.1744548575646958e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50185, + "real_time": 1.3927761741778591e+04, + "cpu_time": 2.1822918262429026e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.1763555626352739e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50185, + "real_time": 1.3936509980626428e+04, + "cpu_time": 2.1889627039952164e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.1756171396408358e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3933977261510585e+04, + "cpu_time": 2.1849821376905442e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.1758313682854478e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3932948791715637e+04, + "cpu_time": 2.1864787187406568e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.1759176212390680e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0576829443807517e+01, + "cpu_time": 4.4452729686914381e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9216975586488028e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5906751140060946e-04, + "cpu_time": 2.0344665029572956e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5875655296201870e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 47768, + "real_time": 1.4713279152326812e+04, + "cpu_time": 2.3045921704906981e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.2271038060756121e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 47768, + "real_time": 1.4673520769289880e+04, + "cpu_time": 2.2999934830849128e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.2331382164653997e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 47768, + "real_time": 1.4702638389535441e+04, + "cpu_time": 2.2999920783788290e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.2287156312925797e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 47768, + "real_time": 1.4712417992904508e+04, + "cpu_time": 2.3075350360073691e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.2272341647581873e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 47768, + "real_time": 1.4717328958990372e+04, + "cpu_time": 2.3075368384692611e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.2264909679811850e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4703837052609404e+04, + "cpu_time": 2.3039299212862141e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.2285365573145928e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4712417992904509e+04, + "cpu_time": 2.3045921704906981e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.2272341647581873e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7785746741946305e+01, + "cpu_time": 3.7897041508548192e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6991096971260728e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2095990099937875e-03, + "cpu_time": 1.6448869020890802e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2111579180816871e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 43396, + "real_time": 1.6063015691105509e+04, + "cpu_time": 2.6595478177712335e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.0799312694619918e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 43396, + "real_time": 1.6137392310054740e+04, + "cpu_time": 2.6738500645220756e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.0611270235505414e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 43396, + "real_time": 1.6297408031159091e+04, + "cpu_time": 2.6915618121485873e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.0212529424741287e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 43396, + "real_time": 1.6248537537733519e+04, + "cpu_time": 2.6766730182505260e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.0333476073035870e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 43396, + "real_time": 1.6245474044348208e+04, + "cpu_time": 2.6824582058254251e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.0341081966026068e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6198365522880214e+04, + "cpu_time": 2.6768181837035700e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.0459534078785715e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6245474044348206e+04, + "cpu_time": 2.6766730182505264e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.0341081966026068e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5633639425650387e+01, + "cpu_time": 1.1792802605565817e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3948183228776250e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.9039067423542037e-03, + "cpu_time": 4.4055299225626253e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9190457265628970e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 35403, + "real_time": 1.9850830167297321e+04, + "cpu_time": 3.3692730164110391e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.6028472812150087e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 35403, + "real_time": 1.9820370071305395e+04, + "cpu_time": 3.3693619919215860e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.6129945873088036e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 35403, + "real_time": 1.9865541154621053e+04, + "cpu_time": 3.3712313560997645e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.5979576886336412e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 35403, + "real_time": 1.9853871420472744e+04, + "cpu_time": 3.3787457503601392e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.6018358447130022e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 35403, + "real_time": 1.9791391646303273e+04, + "cpu_time": 3.3817847809507781e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.6226772903300228e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9836400891999958e+04, + "cpu_time": 3.3740793791486612e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.6076625384400959e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9850830167297321e+04, + "cpu_time": 3.3712313560997645e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.6028472812150087e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0188564334631032e+01, + "cpu_time": 5.8011709491486080e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0064724178714486e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5218771035629809e-03, + "cpu_time": 1.7193344605343412e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5231897997458139e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 29531, + "real_time": 2.3715771694667794e+04, + "cpu_time": 4.9229558599437747e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1053572423238497e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 29531, + "real_time": 2.3681637454959557e+04, + "cpu_time": 4.9082820933933712e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1069504821977592e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 29531, + "real_time": 2.3726357927807276e+04, + "cpu_time": 4.9274252141817109e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1048640537145714e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 29531, + "real_time": 2.3683647536865563e+04, + "cpu_time": 4.9068637364126211e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1068565329388184e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 29531, + "real_time": 2.3786785024043340e+04, + "cpu_time": 4.9355182824828014e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1020572966671562e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3718839927668705e+04, + "cpu_time": 4.9202090372828556e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1052171215684311e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3715771694667790e+04, + "cpu_time": 4.9229558599437747e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1053572423238497e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2734642993478417e+01, + "cpu_time": 1.2392944345351073e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9887805845295250e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8017172477152743e-03, + "cpu_time": 2.5187841108870791e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7994478602604482e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 20412, + "real_time": 3.4267500212328385e+04, + "cpu_time": 7.4300707378013030e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5299861289892908e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 20412, + "real_time": 3.4279445525365751e+04, + "cpu_time": 7.4268317607289660e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5294529767467878e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 20412, + "real_time": 3.4264255161431895e+04, + "cpu_time": 7.4306371105231694e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5301310287641758e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 20412, + "real_time": 3.4245032612889489e+04, + "cpu_time": 7.4113148001175810e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5309899275805134e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 20412, + "real_time": 3.4308564820796964e+04, + "cpu_time": 7.4029835243974609e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5281548579443644e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4272959666562492e+04, + "cpu_time": 7.4203675867136961e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5297429840050266e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4267500212328378e+04, + "cpu_time": 7.4268317607289660e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5299861289892908e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3426948954145054e+01, + "cpu_time": 1.2505582490458251e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0453247213577224e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.8354029479983716e-04, + "cpu_time": 1.6853049858136038e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8333356144635051e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 12719, + "real_time": 5.5171494553180142e+04, + "cpu_time": 1.2479637911785512e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9005756659161575e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 12719, + "real_time": 5.5151418972509578e+04, + "cpu_time": 1.2540827234845480e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9012674914541481e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 12719, + "real_time": 5.5150297513127822e+04, + "cpu_time": 1.2474535482349228e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9013061529729733e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 12719, + "real_time": 5.5181119187223529e+04, + "cpu_time": 1.2483038823806899e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9002441694636456e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 12719, + "real_time": 5.5149784021529165e+04, + "cpu_time": 1.2488452857929088e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9013238557573696e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.5160822849514057e+04, + "cpu_time": 1.2493298462143241e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9009434671128590e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.5151418972509578e+04, + "cpu_time": 1.2483038823806899e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9012674914541481e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4550755245718346e+01, + "cpu_time": 2.7048032924424319e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0137891160757849e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6378785692546161e-04, + "cpu_time": 2.1650033421025143e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6375266823114399e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 7181, + "real_time": 9.7565642293335637e+04, + "cpu_time": 2.2392557095112102e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1494779829305229e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 7181, + "real_time": 9.7575823942770905e+04, + "cpu_time": 2.2347745690015418e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1492536934456207e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 7181, + "real_time": 9.7522483876223632e+04, + "cpu_time": 2.2355816404400382e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1504292309266068e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 7181, + "real_time": 9.7560294420920589e+04, + "cpu_time": 2.2424006572900678e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1495958088768250e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 7181, + "real_time": 9.7555193939170233e+04, + "cpu_time": 2.2357782203035848e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1497081962726273e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7555887694484220e+04, + "cpu_time": 2.2375581593092886e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1496929824904404e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7560294420920603e+04, + "cpu_time": 2.2357782203035848e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1495958088768250e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0176234793584037e+01, + "cpu_time": 3.2075944293283521e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4467280414435063e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0681719238484049e-04, + "cpu_time": 1.4335244945402911e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0685409859281061e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3808, + "real_time": 1.8385272641498581e+05, + "cpu_time": 4.2459245693277224e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2813390270497086e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3808, + "real_time": 1.8385274350049341e+05, + "cpu_time": 4.1620648319327924e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2813388150439774e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3808, + "real_time": 1.8385856514168682e+05, + "cpu_time": 4.2470495509453508e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2812665794317204e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3808, + "real_time": 1.8382621752576949e+05, + "cpu_time": 4.1563224947478774e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2816680103924927e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3808, + "real_time": 1.8382201235090944e+05, + "cpu_time": 4.2578906118697277e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2817202066057404e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8384245298676903e+05, + "cpu_time": 4.2138504117646941e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2814665277047279e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8385272641498581e+05, + "cpu_time": 4.2459245693277230e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2813390270497086e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6973871944690927e+01, + "cpu_time": 5.0153965859726832e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1065048706138800e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.2328358705660445e-05, + "cpu_time": 1.1902170451922411e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.2331175804412593e-05, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1994, + "real_time": 3.5089932637181279e+05, + "cpu_time": 8.1480700752256857e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3906024804138367e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1994, + "real_time": 3.5102634837593109e+05, + "cpu_time": 8.1728673971916211e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3897374196583767e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1994, + "real_time": 3.5089523129939561e+05, + "cpu_time": 8.1673234453359561e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3906303795968540e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1994, + "real_time": 3.5099496818514244e+05, + "cpu_time": 8.1593716700099921e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3899510706305016e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1994, + "real_time": 3.5088624672507891e+05, + "cpu_time": 8.1892273319959780e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3906915925868462e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5094042419147218e+05, + "cpu_time": 8.1673719839518471e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3903225885772835e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5089932637181273e+05, + "cpu_time": 8.1673234453359561e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3906024804138367e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.5239249956212092e+01, + "cpu_time": 1.5365764417067453e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4432166003290908e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8589836182741302e-04, + "cpu_time": 1.8813596891704947e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8588355486251281e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1024, + "real_time": 6.8428063701730932e+05, + "cpu_time": 1.7589919580078134e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4518034111164845e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1024, + "real_time": 6.8431499607868318e+05, + "cpu_time": 1.7588631708984338e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4516803074809338e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1024, + "real_time": 6.8368234212812246e+05, + "cpu_time": 1.7557358095703078e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4539490003174519e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1024, + "real_time": 6.8324889423365681e+05, + "cpu_time": 1.7558527968749988e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4555057668724957e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1024, + "real_time": 6.8424313536752376e+05, + "cpu_time": 1.7588978066406248e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4519377883109554e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8395400096505915e+05, + "cpu_time": 1.7576683083984356e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4529752548196644e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8424313536752376e+05, + "cpu_time": 1.7588631708984338e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4519377883109554e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7211791728945911e+02, + "cpu_time": 1.7118735660304369e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6939683123288937e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.9027729441351418e-04, + "cpu_time": 9.7394574269264365e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9057700806420463e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 519, + "real_time": 1.3493188591368913e+06, + "cpu_time": 3.1425413988439352e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4867681773501270e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 519, + "real_time": 1.3485025562460828e+06, + "cpu_time": 3.4782127167630251e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4882735182503273e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 519, + "real_time": 1.3490062963346697e+06, + "cpu_time": 3.1267186512523978e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4873443579299358e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 519, + "real_time": 1.3496996599351827e+06, + "cpu_time": 3.1491206011560871e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4860665669584152e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 519, + "real_time": 1.3486672179802163e+06, + "cpu_time": 3.1456089826589809e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4879697194873329e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3490389179266084e+06, + "cpu_time": 3.2084404701348851e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4872844679952278e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3490062963346695e+06, + "cpu_time": 3.1456089826589804e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4873443579299358e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8524631705722760e+02, + "cpu_time": 1.5105036494772075e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9459708592058364e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.5969778974428843e-04, + "cpu_time": 4.7079061105774700e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5966818328650457e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 261, + "real_time": 2.6828784734667260e+06, + "cpu_time": 6.9558180766284186e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5013754690604439e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 261, + "real_time": 2.6820820570705724e+06, + "cpu_time": 6.5187433333333582e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5021182265130898e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 261, + "real_time": 2.6832552906839442e+06, + "cpu_time": 6.9365566551724263e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5010241937469315e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 261, + "real_time": 2.6817854112049858e+06, + "cpu_time": 6.9128903793103313e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5023949984814964e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 261, + "real_time": 2.6824304234536216e+06, + "cpu_time": 6.5079811915708389e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5017932772175140e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6824863311759699e+06, + "cpu_time": 6.7663979272030741e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5017412330038956e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6824304234536211e+06, + "cpu_time": 6.9128903793103304e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5017932772175140e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.9222595493370068e+02, + "cpu_time": 2.3151999446941982e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5230665253860559e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2077501311034675e-04, + "cpu_time": 3.4216136408211487e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2076889697958044e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 131, + "real_time": 5.3471960340633644e+06, + "cpu_time": 1.3903034984732741e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5100581154120731e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 131, + "real_time": 5.3473072275803741e+06, + "cpu_time": 1.3278817999999940e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5100059205076336e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 131, + "real_time": 5.3471900942159519e+06, + "cpu_time": 1.3882453244274702e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5100609036731857e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 131, + "real_time": 5.3479654060406536e+06, + "cpu_time": 1.3921216961832045e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5096970120337334e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 131, + "real_time": 5.3474978607801991e+06, + "cpu_time": 1.3167055969465556e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5099164411898926e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3474313245361093e+06, + "cpu_time": 1.3630515832060998e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5099476785633041e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3473072275803741e+06, + "cpu_time": 1.3882453244274700e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5100059205076336e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2350573926383538e+02, + "cpu_time": 3.7441043802535150e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5183693215025126e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.0497408873577181e-05, + "cpu_time": 2.7468545038088911e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0494062663952749e-05, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 66, + "real_time": 1.0677597457260797e+07, + "cpu_time": 2.7944244151515130e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5140061429967388e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 66, + "real_time": 1.0677306616509503e+07, + "cpu_time": 2.7815453909090936e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5140746223859375e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 66, + "real_time": 1.0677420505971620e+07, + "cpu_time": 2.7772871893939547e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5140478063018181e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 66, + "real_time": 1.0677105591942867e+07, + "cpu_time": 2.7833696757576056e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5141219564463814e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 66, + "real_time": 1.0677377001918627e+07, + "cpu_time": 2.6676422984848637e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5140580495730797e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0677361434720684e+07, + "cpu_time": 2.7608537939394064e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5140617155407913e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0677377001918627e+07, + "cpu_time": 2.7815453909090936e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5140580495730797e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7881474499117590e+02, + "cpu_time": 5.2489363443478581e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2103323118252790e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6747091131494851e-05, + "cpu_time": 1.9012004025241253e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6747131885422346e-05, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1336098286238585e+07, + "cpu_time": 5.4960439393939301e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5162562751516407e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1336389028213240e+07, + "cpu_time": 5.5733432848485060e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5162219871886112e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1337067821260653e+07, + "cpu_time": 5.5899188363636911e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5161419389830677e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1337029383038029e+07, + "cpu_time": 5.5846746333333105e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5161464717613785e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1336370063098993e+07, + "cpu_time": 5.6326898606060512e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5162242237657475e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1336590916369904e+07, + "cpu_time": 5.5753341109090984e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5161981793700893e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1336389028213240e+07, + "cpu_time": 5.5846746333333112e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5162219871886112e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3356394490603043e+02, + "cpu_time": 4.9700443374124647e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1129471423045243e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0320207038013305e-05, + "cpu_time": 8.9143434968098484e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0320128931913108e-05, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 16, + "real_time": 4.2656070319935679e+07, + "cpu_time": 1.1211159881250055e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5172075532194008e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 16, + "real_time": 4.2656330624595284e+07, + "cpu_time": 1.1260685224999990e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5171921922906548e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 16, + "real_time": 4.2656330391764641e+07, + "cpu_time": 1.1256154099999982e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5171922060302212e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 16, + "real_time": 4.2655880562961102e+07, + "cpu_time": 1.1248907443749978e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5172187511523327e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 16, + "real_time": 4.2655840516090393e+07, + "cpu_time": 1.1233793468749909e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5172211144098057e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2656090483069427e+07, + "cpu_time": 1.1242140023749983e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5172063634204834e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2656070319935679e+07, + "cpu_time": 1.1248907443749979e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5172075532194008e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3567878658899653e+02, + "cpu_time": 2.0096958760265872e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3907775781914231e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_GPUToPinned/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.5250911164148871e-06, + "cpu_time": 1.7876452986539332e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5250836737182621e-06, + "hip_id": NaN, + "numa_id": NaN + } + ] +} diff --git a/results/hipMemcpyAsync_PageableToGPU.json b/results/hipMemcpyAsync_PageableToGPU.json new file mode 100644 index 0000000..4eebe6b --- /dev/null +++ b/results/hipMemcpyAsync_PageableToGPU.json @@ -0,0 +1,3858 @@ +{ + "context": { + "date": "2023-10-09T17:44:34-04:00", + "host_name": "frontier07674", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.02,0.44,3.92], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 101169, + "real_time": 6.9087199292106898e+03, + "cpu_time": 7.1686062181489611e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.7054621206688225e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 101169, + "real_time": 6.9179356004195024e+03, + "cpu_time": 7.2074324509496796e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.7005259196757510e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 101169, + "real_time": 6.9111886833702492e+03, + "cpu_time": 7.1817291523182774e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.7041384880142108e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 101169, + "real_time": 6.9172644538313225e+03, + "cpu_time": 7.2122679208953989e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.7008849626705706e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 101169, + "real_time": 6.9084828693748696e+03, + "cpu_time": 7.1796930812924778e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.7055892710517026e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9127183072413272e+03, + "cpu_time": 7.1899457647209585e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.7033201524162114e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9111886833702492e+03, + "cpu_time": 7.1817291523182774e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 3.7041384880142108e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5867555773424371e+00, + "cpu_time": 1.8921345990393695e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4568519618609502e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.6352415554639941e-04, + "cpu_time": 2.6316395991796511e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6341873258189416e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 101579, + "real_time": 6.8852134695496316e+03, + "cpu_time": 7.1854867169074619e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.4362255036006942e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 101579, + "real_time": 6.8909391940087180e+03, + "cpu_time": 7.2204639472816110e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.4300466973377883e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 101579, + "real_time": 6.8736316080501674e+03, + "cpu_time": 7.1872791888786280e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.4487553188093856e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 101579, + "real_time": 6.8780371566070735e+03, + "cpu_time": 7.2184497000117195e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.4439842115154967e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 101579, + "real_time": 6.8884894737497016e+03, + "cpu_time": 7.1898920665455163e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.4326890089779913e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8832621803930597e+03, + "cpu_time": 7.2003143239249873e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.4383401480482712e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8852134695496325e+03, + "cpu_time": 7.1898920665455163e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 7.4362255036006942e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2477144759758598e+00, + "cpu_time": 1.7559152254647571e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8345333787341791e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0529476120524572e-03, + "cpu_time": 2.4386646838878336e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0532636613545918e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 100939, + "real_time": 6.9452478403941923e+03, + "cpu_time": 7.1819861047292507e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4743894293365932e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 100939, + "real_time": 6.9320405799333557e+03, + "cpu_time": 7.2021819844675887e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4771985077009526e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 100939, + "real_time": 6.9462324108070161e+03, + "cpu_time": 7.1752828335482228e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4741804469525823e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 100939, + "real_time": 6.9321389351595999e+03, + "cpu_time": 7.2135237164621961e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4771775487740198e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 100939, + "real_time": 6.9418255930113364e+03, + "cpu_time": 7.1819711197175336e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4751162878982571e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9394970718611003e+03, + "cpu_time": 7.1909891517849565e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4756124441324809e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9418255930113355e+03, + "cpu_time": 7.1819861047292495e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.4751162878982571e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9570034447699003e+00, + "cpu_time": 1.6141207358387648e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4796707662517362e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0025227149356090e-03, + "cpu_time": 2.2446435417554558e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0027502628724720e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 99950, + "real_time": 7.0037144331567461e+03, + "cpu_time": 7.1380743445125651e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.9241626276257467e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 99950, + "real_time": 7.0068612133665320e+03, + "cpu_time": 7.1713721193240094e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.9228493866742557e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 99950, + "real_time": 7.0075833214079912e+03, + "cpu_time": 7.1268622173888481e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.9225481968133169e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 99950, + "real_time": 7.0077959976445036e+03, + "cpu_time": 7.1714443136720288e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.9224595018011147e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 99950, + "real_time": 7.0046412959360541e+03, + "cpu_time": 7.1361899109959543e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.9237756988187343e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0061192523023665e+03, + "cpu_time": 7.1487885811786830e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.9231590823466337e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0068612133665329e+03, + "cpu_time": 7.1380743445125651e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.9228493866742557e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8352785671034304e+00, + "cpu_time": 2.1080709170842674e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6580423085799150e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6195365808258505e-04, + "cpu_time": 2.9488505543923795e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6197829447011296e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 96223, + "real_time": 7.2655943578245642e+03, + "cpu_time": 7.0149601862444973e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.6375291521592844e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 96223, + "real_time": 7.2702245780778567e+03, + "cpu_time": 7.0579324229389930e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.6339387539014971e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 96223, + "real_time": 7.2669438825198404e+03, + "cpu_time": 7.0132668846033621e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.6364822217117441e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 96223, + "real_time": 7.2722469359854649e+03, + "cpu_time": 7.0559381479271853e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.6323719973419046e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 96223, + "real_time": 7.2669623437887276e+03, + "cpu_time": 7.0161693388376689e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.6364679025768781e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2683944196392913e+03, + "cpu_time": 7.0316533961103415e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.6353580055382621e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2669623437887276e+03, + "cpu_time": 7.0161693388376689e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.6364679025768781e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7471170087502452e+00, + "cpu_time": 2.3112872563261278e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1295616469123404e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7795376119483843e-04, + "cpu_time": 3.2869755178840995e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7789287651635809e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 76701, + "real_time": 9.1190849191778980e+03, + "cpu_time": 6.3122155873929012e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 8.9833575107649326e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 76701, + "real_time": 9.0900220646406688e+03, + "cpu_time": 6.3260775981121409e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 9.0120793346213198e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 76701, + "real_time": 9.1275763475916865e+03, + "cpu_time": 6.3094900651761878e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 8.9750002498324323e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 76701, + "real_time": 9.1119896391112979e+03, + "cpu_time": 6.3479176042866266e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 8.9903526281873333e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 76701, + "real_time": 9.1155714058371432e+03, + "cpu_time": 6.3100102269861901e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 8.9868200634732175e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1128488752717403e+03, + "cpu_time": 6.3211422163908076e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 8.9895219573758471e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1155714058371414e+03, + "cpu_time": 6.3122155873929000e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 8.9868200634732175e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4008929540950732e+01, + "cpu_time": 1.6436680861813850e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3835695726055845e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5372722331613332e-03, + "cpu_time": 2.6002706946211891e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5390913767893677e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 71438, + "real_time": 9.8034233746711780e+03, + "cpu_time": 6.1232005741586471e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.6712529260269291e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 71438, + "real_time": 9.8040390501506445e+03, + "cpu_time": 6.1867302599805188e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.6711479744410288e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 71438, + "real_time": 9.7998460757164066e+03, + "cpu_time": 6.1233201909945297e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.6718629939095514e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 71438, + "real_time": 9.8044724334979783e+03, + "cpu_time": 6.1992008157375908e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.6710741053258915e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 71438, + "real_time": 9.8022094717502532e+03, + "cpu_time": 6.1098116466394401e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.6714598935289357e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8027980811572925e+03, + "cpu_time": 6.1484526975021470e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.6713595786464672e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8034233746711780e+03, + "cpu_time": 6.1233201909945285e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.6712529260269291e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8570977919849914e+00, + "cpu_time": 4.1240110093508684e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1667654766338476e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8944568444744983e-04, + "cpu_time": 6.7073964983519795e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8947242215815814e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 65430, + "real_time": 1.0700698275981844e+04, + "cpu_time": 5.9734941192293406e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0622300671302094e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 65430, + "real_time": 1.0702140803440796e+04, + "cpu_time": 6.0064937303895438e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0618173131739130e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 65430, + "real_time": 1.0701547515220947e+04, + "cpu_time": 5.9711721267965400e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0619870587308669e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 65430, + "real_time": 1.0697345561584942e+04, + "cpu_time": 6.0028608106639254e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0631898176378093e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 65430, + "real_time": 1.0697074575490955e+04, + "cpu_time": 5.9652560726162934e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0632674165960999e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0699761346343896e+04, + "cpu_time": 5.9838553719391286e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0624983346537800e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0700698275981844e+04, + "cpu_time": 5.9734941192293406e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.0622300671302094e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3866716437228095e+00, + "cpu_time": 1.9286369375838016e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8315381389552378e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2305839975938661e-04, + "cpu_time": 3.2230674334610583e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2307075441161189e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56332, + "real_time": 1.2421327900816234e+04, + "cpu_time": 5.7192233598734081e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2760864638066177e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56332, + "real_time": 1.2445777529726538e+04, + "cpu_time": 5.7495669139682949e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2657216347848358e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56332, + "real_time": 1.2418913886451544e+04, + "cpu_time": 5.7050918892015767e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2771120404898462e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56332, + "real_time": 1.2413332897985680e+04, + "cpu_time": 5.7382799073911762e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2794846105057392e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56332, + "real_time": 1.2422504305460283e+04, + "cpu_time": 5.7206057205944490e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2755868211849842e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2424371304088056e+04, + "cpu_time": 5.7265535582057810e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2747983141544037e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2421327900816232e+04, + "cpu_time": 5.7206057205944479e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.2760864638066177e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2476019997832688e+01, + "cpu_time": 1.7441571208778075e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2910746629243474e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0041570468622132e-03, + "cpu_time": 3.0457361537788172e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0030856817266867e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 45316, + "real_time": 1.5461537606810714e+04, + "cpu_time": 5.4325483017397785e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4772940009707422e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 45316, + "real_time": 1.5467105217171222e+04, + "cpu_time": 5.4679050738730085e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4742424752168179e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 45316, + "real_time": 1.5462244022641009e+04, + "cpu_time": 5.4352802101574135e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4769067030680847e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 45316, + "real_time": 1.5452947006917910e+04, + "cpu_time": 5.4619204783745039e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4820066969311581e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 45316, + "real_time": 1.5468288863693235e+04, + "cpu_time": 5.4320249121425653e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4735940190287485e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5462424543446821e+04, + "cpu_time": 5.4459357952574527e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4768087790431108e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5462244022641011e+04, + "cpu_time": 5.4352802101574135e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4769067030680847e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.0609797328651087e+00, + "cpu_time": 1.7495988417514805e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3235514529611240e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.9198120035022790e-04, + "cpu_time": 3.2126688736857740e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9207578460160773e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 36669, + "real_time": 1.9099041969663256e+04, + "cpu_time": 5.4591137456954896e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3725505206825930e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 36669, + "real_time": 1.9091606669986122e+04, + "cpu_time": 5.4760579206417501e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3730850657640881e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 36669, + "real_time": 1.9096898418608940e+04, + "cpu_time": 5.4575755285227799e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3727045840310604e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 36669, + "real_time": 1.9093777880056874e+04, + "cpu_time": 5.4770907472696197e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3729289281918636e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 36669, + "real_time": 1.9099358943634539e+04, + "cpu_time": 5.4586724952445042e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3725277417615511e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9096136776389947e+04, + "cpu_time": 5.4657020874748302e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3727593680862314e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9096898418608940e+04, + "cpu_time": 5.4591137456954896e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3727045840310604e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3722716466749736e+00, + "cpu_time": 9.9474522769451630e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4243498540928452e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7659444348159349e-04, + "cpu_time": 1.8199770345589609e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7660413838389169e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 22742, + "real_time": 3.0771388682977416e+04, + "cpu_time": 5.3503187310584021e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7038165076054354e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 22742, + "real_time": 3.0776346378508646e+04, + "cpu_time": 5.3771839199059665e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7035420434639837e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 22742, + "real_time": 3.0774348296952903e+04, + "cpu_time": 5.4211203281526864e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7036526490860312e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 22742, + "real_time": 3.0778563686839672e+04, + "cpu_time": 5.3844742923024070e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7034193191548298e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 22742, + "real_time": 3.0773834181777471e+04, + "cpu_time": 5.3717516396122217e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7036811107224781e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0774896245411226e+04, + "cpu_time": 5.3809697822063363e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7036223260065516e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0774348296952900e+04, + "cpu_time": 5.3771839199059653e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7036526490860312e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7057454395772362e+00, + "cpu_time": 2.5802818515950413e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4978208931110555e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.7920538155532668e-05, + "cpu_time": 4.7951985534790704e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7919773663807657e-05, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 13274, + "real_time": 5.2816056328625884e+04, + "cpu_time": 5.2642594728492332e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9853356590573009e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 13274, + "real_time": 5.2841008208197331e+04, + "cpu_time": 5.2701232761428583e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9843981702024609e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 13274, + "real_time": 5.2815479039162732e+04, + "cpu_time": 5.2643943923270416e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9853573593879169e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 13274, + "real_time": 5.2848854462155236e+04, + "cpu_time": 5.2730599583410591e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9841035546964962e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 13274, + "real_time": 5.2824905211900012e+04, + "cpu_time": 5.2760712386395109e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9850030885881920e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2829260650008240e+04, + "cpu_time": 5.2695816676599419e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9848395663864735e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2824905211900012e+04, + "cpu_time": 5.2701232761428583e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9850030885881920e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5043099239414602e+01, + "cpu_time": 5.2378421190530929e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6513291364633860e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8474938044419242e-04, + "cpu_time": 9.9397683713649634e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8472473202214470e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 7213, + "real_time": 9.7084159462229276e+04, + "cpu_time": 4.9987841770553732e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1601381848661930e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 7213, + "real_time": 9.7511807260330184e+04, + "cpu_time": 5.0110498810799330e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1506646824842152e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 7213, + "real_time": 9.6897394892416938e+04, + "cpu_time": 4.9824255802107376e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1643017362111973e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 7213, + "real_time": 9.6969722918429470e+04, + "cpu_time": 4.9908672348732132e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1626874212728394e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 7213, + "real_time": 9.7556538756298105e+04, + "cpu_time": 5.0067416059878439e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1496785625398285e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7203924657940806e+04, + "cpu_time": 4.9979736958414209e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1574941174748550e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7084159462229305e+04, + "cpu_time": 4.9987841770553732e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1601381848661930e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0914521396432025e+02, + "cpu_time": 1.1627181899264180e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8561879130149871e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1803779019437509e-03, + "cpu_time": 2.3263791702102417e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1778477899348867e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3779, + "real_time": 1.8501800339428455e+05, + "cpu_time": 4.8321096079412103e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2669707396321232e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3779, + "real_time": 1.8500094533427592e+05, + "cpu_time": 4.8304670143662363e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2671797662554447e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3779, + "real_time": 1.8506487625305174e+05, + "cpu_time": 4.8339816403360772e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2663965658533951e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3779, + "real_time": 1.8507477923761684e+05, + "cpu_time": 4.8301878228790885e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2662752954658119e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3779, + "real_time": 1.8498972423268558e+05, + "cpu_time": 4.8294551703995192e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2673172887831757e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8502966569038294e+05, + "cpu_time": 4.8312402511844271e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2668279311979904e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8501800339428458e+05, + "cpu_time": 4.8304670143662369e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.2669707396321232e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8181159924952127e+01, + "cpu_time": 1.8140431075447029e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6774333512985520e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0635155872162728e-04, + "cpu_time": 3.7548186660764225e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0634267325383567e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1932, + "real_time": 3.6198252604565222e+05, + "cpu_time": 4.7729859675828022e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3174068902270859e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1932, + "real_time": 3.6176082365529123e+05, + "cpu_time": 4.7689466196635735e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3188270955489643e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1932, + "real_time": 3.6172621206049743e+05, + "cpu_time": 4.7666014069307107e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3190489713797783e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1932, + "real_time": 3.6214070272796805e+05, + "cpu_time": 4.7730088593633848e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3163946877027336e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1932, + "real_time": 3.6146475852266588e+05, + "cpu_time": 4.7661145999792600e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3207263784953430e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6181500460241496e+05, + "cpu_time": 4.7695314907039464e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3184808046707813e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6176082365529123e+05, + "cpu_time": 4.7689466196635735e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.3188270955489643e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5866711704306539e+02, + "cpu_time": 3.3402137121802371e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6575968158116134e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.1491539530624238e-04, + "cpu_time": 7.0032323273061092e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1494955337660776e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 980, + "real_time": 7.1489371689587680e+05, + "cpu_time": 4.7434996626428938e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.3468126245182228e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 980, + "real_time": 7.1446727117884671e+05, + "cpu_time": 4.7428779007550889e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.3482133719460884e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 980, + "real_time": 7.1466400190636679e+05, + "cpu_time": 4.7415031305509925e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.3475669622713284e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 980, + "real_time": 7.1403575343631057e+05, + "cpu_time": 4.7413078745101613e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.3496324825835865e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 980, + "real_time": 7.1461666137816350e+05, + "cpu_time": 4.7442856863877583e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.3477224793002373e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1453548095911299e+05, + "cpu_time": 4.7426948509693795e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.3479895841238930e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1461666137816338e+05, + "cpu_time": 4.7428779007550889e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.3477224793002373e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1852042764512169e+02, + "cpu_time": 1.2802233794465009e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0469460902629131e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.4577272386470628e-04, + "cpu_time": 2.6993585285901974e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4589043211346307e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 71, + "real_time": 9.7800851471617185e+06, + "cpu_time": 7.0437927447887731e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4308936471517162e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 71, + "real_time": 9.7798192611252759e+06, + "cpu_time": 7.0293812215493405e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4309869235905690e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 71, + "real_time": 9.7577843920026030e+06, + "cpu_time": 7.0216311325352132e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4387347221466513e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 71, + "real_time": 9.7654419754382595e+06, + "cpu_time": 7.0281854912675858e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4360382340497317e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 71, + "real_time": 9.7737756990630869e+06, + "cpu_time": 7.0205901384506869e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4331084560510764e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7713812949581891e+06, + "cpu_time": 7.0287161457183206e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4339523965979490e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7737756990630869e+06, + "cpu_time": 7.0281854912675858e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4331084560510764e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6545426303718232e+03, + "cpu_time": 9.2770842057368497e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3942440067856051e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.8804276887172015e-04, + "cpu_time": 1.3198831782939714e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.8843653457407173e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 37, + "real_time": 1.8907663739613585e+07, + "cpu_time": 7.3318513616216111e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5492943456255641e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 37, + "real_time": 1.8874163160452973e+07, + "cpu_time": 7.3147882154054666e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5555941436711316e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 37, + "real_time": 1.8921307061572332e+07, + "cpu_time": 7.3309325445945370e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5467351056467319e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 37, + "real_time": 1.8945960025932338e+07, + "cpu_time": 7.3333895551352119e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5421200038501372e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 37, + "real_time": 1.8880087831938587e+07, + "cpu_time": 7.3240406805405164e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5544783794106603e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8905836363901965e+07, + "cpu_time": 7.3270004714594686e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5496443956408458e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8907663739613585e+07, + "cpu_time": 7.3309325445945370e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5492943456255641e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9659605348964484e+04, + "cpu_time": 7.7089604514000833e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5670880131285870e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5688068371095885e-03, + "cpu_time": 1.0521304702283624e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5683509086051749e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 19, + "real_time": 3.7375963832202710e+07, + "cpu_time": 7.8848779789473534e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5910171735654216e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 19, + "real_time": 3.7407627427264266e+07, + "cpu_time": 7.8891610057895434e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5879775658313050e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 19, + "real_time": 3.7297791556308143e+07, + "cpu_time": 7.8728007273683512e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5985435705321236e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 19, + "real_time": 3.7320940706290700e+07, + "cpu_time": 7.8761884494736922e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5963114932249470e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 19, + "real_time": 3.7243821511143133e+07, + "cpu_time": 7.8753096878948796e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.6037582222823949e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7329229006641790e+07, + "cpu_time": 7.8796675698947644e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5955216050872383e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7320940706290692e+07, + "cpu_time": 7.8761884494736922e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5963114932249470e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4579226086062183e+04, + "cpu_time": 6.9914703084871219e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2209423096891036e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7299908892994265e-03, + "cpu_time": 8.8727985622120552e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7301918867313173e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10, + "real_time": 7.4604365974664688e+07, + "cpu_time": 9.1785468359998667e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5981199289483876e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10, + "real_time": 7.4093629419803619e+07, + "cpu_time": 9.1492674490000081e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6229222147977681e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10, + "real_time": 7.4396270513534546e+07, + "cpu_time": 9.1730972780000567e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6081843101417947e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10, + "real_time": 7.4398460984230042e+07, + "cpu_time": 9.1759716240001416e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6080780764658456e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10, + "real_time": 7.4280925095081329e+07, + "cpu_time": 9.1649219609998631e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6137871957894483e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4354730397462860e+07, + "cpu_time": 9.1683610295999873e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6102183452286491e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4396270513534546e+07, + "cpu_time": 9.1730972780000567e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6081843101417947e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8679989966156034e+05, + "cpu_time": 1.1837396894207706e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.0721375920430124e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5122799674347869e-03, + "cpu_time": 1.2911137395212465e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5129055155439467e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5, + "real_time": 1.4645697176456451e+08, + "cpu_time": 1.1019845420000081e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6657245164336843e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5, + "real_time": 1.4515162110328674e+08, + "cpu_time": 1.0957013120000169e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6986904308700371e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5, + "real_time": 1.4613095223903656e+08, + "cpu_time": 1.0994548989999998e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6739027822237339e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5, + "real_time": 1.4453053176403046e+08, + "cpu_time": 1.0936148306000121e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7145847693726668e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5, + "real_time": 1.4586304724216461e+08, + "cpu_time": 1.0958107964000192e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6806505976025352e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4562662482261661e+08, + "cpu_time": 1.0973132760000112e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6867106193005314e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4586304724216461e+08, + "cpu_time": 1.0958107964000192e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6806505976025352e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7890408168567182e+05, + "cpu_time": 3.3519265497423462e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9758954897706509e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3486378787837149e-03, + "cpu_time": 3.0546669060279481e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3595079565684264e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 2.8557619452476501e+08, + "cpu_time": 1.3170925220000100e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7599136223060975e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 2.8420419991016388e+08, + "cpu_time": 1.3143135904999924e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7780645899652667e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 2.8506968915462494e+08, + "cpu_time": 1.3163117245000110e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7665941517114105e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 2.8370851278305054e+08, + "cpu_time": 1.3139500370000262e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7846655127372971e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 2.8643620014190674e+08, + "cpu_time": 1.3200945764999971e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7486247320277424e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8499895930290228e+08, + "cpu_time": 1.3163524901000075e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7675725217495632e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8506968915462494e+08, + "cpu_time": 1.3163117245000110e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7665941517114105e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0840617089758967e+06, + "cpu_time": 2.4739347277471600e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4325576081825122e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PageableToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.8037391842674611e-03, + "cpu_time": 1.8793862178657081e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8023358539552932e-03, + "hip_id": NaN, + "numa_id": NaN + } + ] +} diff --git a/results/hipMemcpyAsync_PinnedToGPU.json b/results/hipMemcpyAsync_PinnedToGPU.json new file mode 100644 index 0000000..a2ecc50 --- /dev/null +++ b/results/hipMemcpyAsync_PinnedToGPU.json @@ -0,0 +1,3858 @@ +{ + "context": { + "date": "2023-10-09T17:41:29-04:00", + "host_name": "frontier08191", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [1.53,7.18,22.42], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53841, + "real_time": 1.2985156955162200e+04, + "cpu_time": 5.4588967601355541e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9714817532353982e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53841, + "real_time": 1.2824764378240932e+04, + "cpu_time": 5.4089799243990910e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9961380377041548e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53841, + "real_time": 1.2807087310059263e+04, + "cpu_time": 5.4026471667929828e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9988932206226632e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53841, + "real_time": 1.2785066891794657e+04, + "cpu_time": 5.3962137713894176e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0023360234767213e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53841, + "real_time": 1.2777480354687712e+04, + "cpu_time": 5.3952083580811727e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 2.0035248960964397e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2835911177988955e+04, + "cpu_time": 5.4123891961596441e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9944747862270754e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2807087310059262e+04, + "cpu_time": 5.4026471667929840e+08, + "time_unit": "ns", + "bytes": 2.5600000000000000e+02, + "bytes_per_second": 1.9988932206226632e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.5481973335537361e+01, + "cpu_time": 2.6581522438943950e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3178316222913068e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):8/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.6595952675429860e-03, + "cpu_time": 4.9112363275362466e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6074117927769523e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54750, + "real_time": 1.2778965156013996e+04, + "cpu_time": 5.4558934066469383e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0065842088867746e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54750, + "real_time": 1.2852974745079360e+04, + "cpu_time": 5.5084637867271066e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9835136235369503e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54750, + "real_time": 1.2785234727191613e+04, + "cpu_time": 5.4768635764889514e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 4.0046194764893860e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54750, + "real_time": 1.2833472019146897e+04, + "cpu_time": 5.4913035341066611e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9895672756064892e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54750, + "real_time": 1.2848108899735762e+04, + "cpu_time": 5.4977309156140685e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9850222627746403e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2819751109433526e+04, + "cpu_time": 5.4860510439167440e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9938613694588482e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2833472019146899e+04, + "cpu_time": 5.4913035341066623e+08, + "time_unit": "ns", + "bytes": 5.1200000000000000e+02, + "bytes_per_second": 3.9895672756064892e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5182070145883998e+01, + "cpu_time": 2.0372970695148886e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0968732887735940e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):9/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7443645235822839e-03, + "cpu_time": 3.7135948120168571e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7463980025982117e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54509, + "real_time": 1.2893073067507357e+04, + "cpu_time": 5.4658834072635770e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9422492577091396e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54509, + "real_time": 1.2812681327612734e+04, + "cpu_time": 5.3987539397438800e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9920820148173645e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54509, + "real_time": 1.2812343455595210e+04, + "cpu_time": 5.4356995696095884e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9922927725826323e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54509, + "real_time": 1.2821466963756589e+04, + "cpu_time": 5.4426244541695929e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9866056114687830e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54509, + "real_time": 1.2849437470461104e+04, + "cpu_time": 5.4478786737652910e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9692204608491197e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2837800456986599e+04, + "cpu_time": 5.4381680089103866e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9764900234854087e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2821466963756588e+04, + "cpu_time": 5.4426244541695929e+08, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 7.9866056114687830e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4410369577830195e+01, + "cpu_time": 2.4707601375766052e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1333300112148613e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6803944876011335e-03, + "cpu_time": 4.5433685269162119e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6745222584540775e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54620, + "real_time": 1.2827654443934500e+04, + "cpu_time": 5.4572774029297614e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5965506468475130e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54620, + "real_time": 1.2840900522921744e+04, + "cpu_time": 5.4642324980884349e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5949037190531948e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54620, + "real_time": 1.2939033648107416e+04, + "cpu_time": 5.5122906789369881e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5828075385672715e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54620, + "real_time": 1.2906854669116152e+04, + "cpu_time": 5.4600892369969046e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5867537463643301e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54620, + "real_time": 1.2914699274505330e+04, + "cpu_time": 5.4648544187676609e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5857899254711404e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2885828511717027e+04, + "cpu_time": 5.4717488471439493e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5893611152606902e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2906854669116150e+04, + "cpu_time": 5.4642324980884337e+08, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.5867537463643301e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8757443320798323e+01, + "cpu_time": 2.2874865995933511e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0189483671983762e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7838035231078379e-03, + "cpu_time": 4.1805401956402557e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7870237980568290e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53998, + "real_time": 1.2926423786894104e+04, + "cpu_time": 5.4418909358818161e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1687031676563704e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53998, + "real_time": 1.2954910975479610e+04, + "cpu_time": 5.4486743489475608e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1617353509821093e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53998, + "real_time": 1.2947645644946700e+04, + "cpu_time": 5.4561148817248654e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1635094999673676e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53998, + "real_time": 1.2937754835712176e+04, + "cpu_time": 5.4403570292553139e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1659279774677616e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53998, + "real_time": 1.2927069403712248e+04, + "cpu_time": 5.4428225639625335e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1685449130672705e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2938760929348968e+04, + "cpu_time": 5.4459719519544184e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1656841818281758e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2937754835712174e+04, + "cpu_time": 5.4428225639625335e+08, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.1659279774677616e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2546625455839797e+01, + "cpu_time": 6.4861904767974070e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0692019004946545e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.6969296552812162e-04, + "cpu_time": 1.1910069559703995e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6952245524447645e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53454, + "real_time": 1.3086313163026587e+04, + "cpu_time": 5.4814156508352840e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2599755163625956e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53454, + "real_time": 1.3083942240193959e+04, + "cpu_time": 5.4715034524479258e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2611098777508509e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53454, + "real_time": 1.3081939648695577e+04, + "cpu_time": 5.4321425158259320e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2620683323644876e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53454, + "real_time": 1.3037909541199251e+04, + "cpu_time": 5.4580269327940238e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2832158591940081e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53454, + "real_time": 1.3111178342972020e+04, + "cpu_time": 5.4743551810507905e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2481035538587987e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3080256587217478e+04, + "cpu_time": 5.4634887465907907e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2628946279061484e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3083942240193959e+04, + "cpu_time": 5.4715034524479258e+08, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.2611098777508509e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6469445260866895e+01, + "cpu_time": 1.9468605476332814e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2689736227991502e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0236181977296868e-03, + "cpu_time": 3.5634017711633791e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0261775076733197e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52056, + "real_time": 1.3429151169470766e+04, + "cpu_time": 5.4129159751649892e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2200324349052424e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52056, + "real_time": 1.3449262886699133e+04, + "cpu_time": 5.4440230337321341e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2182080265680003e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52056, + "real_time": 1.3505481960299852e+04, + "cpu_time": 5.4142814066553330e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2131370097092216e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52056, + "real_time": 1.3443598011224904e+04, + "cpu_time": 5.4177499390185487e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2187213561667025e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52056, + "real_time": 1.3432520487515963e+04, + "cpu_time": 5.4150702091696680e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2197264106335895e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3452002903042125e+04, + "cpu_time": 5.4208081127481353e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2179650475965514e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3443598011224904e+04, + "cpu_time": 5.4150702091696668e+08, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2187213561667025e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0983567027428812e+01, + "cpu_time": 1.3096681338855510e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7981012002284694e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3032679408968894e-03, + "cpu_time": 2.4160016489157762e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2973575520496671e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49753, + "real_time": 1.4088055143683552e+04, + "cpu_time": 5.3516815795112044e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3259420598372436e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49753, + "real_time": 1.4077257312186339e+04, + "cpu_time": 5.3728052406112289e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3277261524255538e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49753, + "real_time": 1.4181585486297872e+04, + "cpu_time": 5.3714564854469073e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3106020149623017e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49753, + "real_time": 1.4069928862580819e+04, + "cpu_time": 5.3542472142442125e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3289385696289463e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49753, + "real_time": 1.4092061435306110e+04, + "cpu_time": 5.3468691447271788e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3252808079521551e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4101777648010937e+04, + "cpu_time": 5.3594119329081470e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3236979209612403e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4088055143683552e+04, + "cpu_time": 5.3542472142442119e+08, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3259420598372436e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5463158075421291e+01, + "cpu_time": 1.1918537608904571e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.4621627501737056e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.2239309972267138e-03, + "cpu_time": 2.2238517505478038e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2113308200950857e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 43073, + "real_time": 1.6604095860073892e+04, + "cpu_time": 5.2501122940238440e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.9469779355819945e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 43073, + "real_time": 1.6408417487541668e+04, + "cpu_time": 5.1729904976744401e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.9940475703863077e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 43073, + "real_time": 1.6472498285102367e+04, + "cpu_time": 5.2031195901811403e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.9785100514638014e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 43073, + "real_time": 1.6297390512990372e+04, + "cpu_time": 5.1549206587542695e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.0212572649444933e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 43073, + "real_time": 1.6456470471927329e+04, + "cpu_time": 5.1801012457774109e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.9823849294899645e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6447774523527125e+04, + "cpu_time": 5.1922488572822207e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.9846355503733120e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6456470471927329e+04, + "cpu_time": 5.1801012457774115e+08, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.9823849294899645e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1100402140415193e+02, + "cpu_time": 3.6668083637980358e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6881308942749050e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.7488778646235831e-03, + "cpu_time": 7.0620813150264793e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7462403030135584e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 35050, + "real_time": 1.9962606268183685e+04, + "cpu_time": 4.9574489365297854e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.5658761305582619e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 35050, + "real_time": 1.9973644356262575e+04, + "cpu_time": 4.9654693608810115e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.5622476130102625e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 35050, + "real_time": 1.9959132511839809e+04, + "cpu_time": 4.9552935619209224e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.5670188783128605e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 35050, + "real_time": 1.9976086854079349e+04, + "cpu_time": 4.9618800625862956e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.5614452398735733e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 35050, + "real_time": 1.9979486718766202e+04, + "cpu_time": 4.9662203889645994e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.5603286933736658e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9970191341826328e+04, + "cpu_time": 4.9612624621765232e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.5633833110257254e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9973644356262575e+04, + "cpu_time": 4.9618800625862962e+08, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.5622476130102625e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8447084255933408e+00, + "cpu_time": 4.8174477974037855e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9072374117570790e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.4289552734873636e-04, + "cpu_time": 9.7101248606193543e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4294798490181978e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 30354, + "real_time": 2.3050518142365567e+04, + "cpu_time": 5.0539941311349273e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1372586003530827e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 30354, + "real_time": 2.3072719641528307e+04, + "cpu_time": 5.0611390431959897e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1361642843705786e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 30354, + "real_time": 2.3077390003130378e+04, + "cpu_time": 5.0686361411619705e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1359343494409071e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 30354, + "real_time": 2.3067917372752145e+04, + "cpu_time": 5.0759217997888505e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1364008105458399e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 30354, + "real_time": 2.3065841368667334e+04, + "cpu_time": 5.0762287718742102e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1365030904794859e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3066877305688748e+04, + "cpu_time": 5.0671839774311906e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1364522270379789e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3067917372752148e+04, + "cpu_time": 5.0686361411619711e+08, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1364008105458399e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0178737948544908e+01, + "cpu_time": 9.6278229729932605e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0164151866447423e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.4127073698157793e-04, + "cpu_time": 1.9000342233230078e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4141012418264192e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 20558, + "real_time": 3.3941833972609093e+04, + "cpu_time": 5.1005205991112596e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5446660908868332e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 20558, + "real_time": 3.4055556972986764e+04, + "cpu_time": 5.0916810926189762e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5395079293986320e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 20558, + "real_time": 3.4075768600924523e+04, + "cpu_time": 5.1074368566300243e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5385947889837923e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 20558, + "real_time": 3.4022736088914295e+04, + "cpu_time": 5.1046461035523504e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5409930542618231e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 20558, + "real_time": 3.4020424373168833e+04, + "cpu_time": 5.1023360856630182e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5410977660040434e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4023264001720701e+04, + "cpu_time": 5.1013241475151253e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5409719259070250e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4022736088914295e+04, + "cpu_time": 5.1023360856630182e+08, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.5409930542618231e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1085126880603781e+01, + "cpu_time": 5.9803599649139040e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3160960020337671e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5014763685818088e-03, + "cpu_time": 1.1723152248278440e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5030098622143941e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 12697, + "real_time": 5.4873434441454912e+04, + "cpu_time": 5.2511963612200427e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9108991640002735e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 12697, + "real_time": 5.4699599745142616e+04, + "cpu_time": 5.2445479970458150e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9169719794761654e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 12697, + "real_time": 5.4970165408640503e+04, + "cpu_time": 5.2617782828117424e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9075365558845112e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 12697, + "real_time": 5.4874255532727278e+04, + "cpu_time": 5.2509237787257057e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9108705709449196e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 12697, + "real_time": 5.4878855325568911e+04, + "cpu_time": 5.2518119646011406e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9107104070945374e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4859262090706849e+04, + "cpu_time": 5.2520516768808901e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9113977354800812e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4874255532727286e+04, + "cpu_time": 5.2511963612200415e+08, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9108705709449196e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8235894911447346e+01, + "cpu_time": 6.1841178656064067e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4270906131114773e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7906893233273821e-03, + "cpu_time": 1.1774670635531631e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7929761815118522e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 7256, + "real_time": 9.6021084474022195e+04, + "cpu_time": 5.0251871095052636e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1840536497662334e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 7256, + "real_time": 9.6101212935675139e+04, + "cpu_time": 5.0304053600454986e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1822326024164944e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 7256, + "real_time": 9.6057863632547756e+04, + "cpu_time": 5.0227487760543340e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1832174073975677e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 7256, + "real_time": 9.5858811420564874e+04, + "cpu_time": 5.0164831092833376e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1877508900033073e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 7256, + "real_time": 9.5841325223738211e+04, + "cpu_time": 5.0200365319087487e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1881500439442715e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5976059537309629e+04, + "cpu_time": 5.0229721773594368e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1850809187055752e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6021084474022195e+04, + "cpu_time": 5.0227487760543334e+08, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.1840536497662334e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1862017553794612e+02, + "cpu_time": 5.2662618642041716e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7013486449371915e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2359350457791388e-03, + "cpu_time": 1.0484354040305739e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2362693856378781e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3852, + "real_time": 1.8124428486314541e+05, + "cpu_time": 4.8797079057866269e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.3141717285966011e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3852, + "real_time": 1.8181475424589752e+05, + "cpu_time": 4.7895734746573144e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.3069106890672710e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3852, + "real_time": 1.8130692185581152e+05, + "cpu_time": 4.8851585556309110e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.3133722403249535e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3852, + "real_time": 1.8178434794347273e+05, + "cpu_time": 4.7932436117808688e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.3072965563043148e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3852, + "real_time": 1.8133039148086790e+05, + "cpu_time": 4.8807804258151472e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.3130728201414265e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8149614007783905e+05, + "cpu_time": 4.8456927947341728e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.3109648068869133e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8133039148086790e+05, + "cpu_time": 4.8797079057866275e+08, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.3130728201414265e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7896502648773946e+02, + "cpu_time": 4.9613533426706027e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5502019705746323e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5370300788110343e-03, + "cpu_time": 1.0238687330864469e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5362423348008868e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2004, + "real_time": 3.4918157270404988e+05, + "cpu_time": 4.7686286538024068e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4023627406907280e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2004, + "real_time": 3.4916040849729726e+05, + "cpu_time": 4.7580495604791087e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4025083588665047e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2004, + "real_time": 3.4926380766829860e+05, + "cpu_time": 4.7708126248652333e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4017970988757000e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2004, + "real_time": 3.4934875641950872e+05, + "cpu_time": 4.7697541141117465e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4012130702783161e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2004, + "real_time": 3.4926875600205112e+05, + "cpu_time": 4.7730084012724781e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4017630709432072e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4924466025824111e+05, + "cpu_time": 4.7680506709061956e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4019288679308914e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4926380766829860e+05, + "cpu_time": 4.7697541141117465e+08, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4017970988757000e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5600326854938785e+01, + "cpu_time": 5.8194817811448197e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1992146429706095e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1646809660321741e-04, + "cpu_time": 1.2205159262786934e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1645997566320112e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1026, + "real_time": 6.8182849896423600e+05, + "cpu_time": 4.8210148291131103e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4606211130051365e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1026, + "real_time": 6.8148510360620101e+05, + "cpu_time": 4.8135477996783257e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4618610019823387e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1026, + "real_time": 6.8180775666994287e+05, + "cpu_time": 4.8149608557992750e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4606959712430645e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1026, + "real_time": 6.8174771711295634e+05, + "cpu_time": 4.8171758551949340e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4609126776467434e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1026, + "real_time": 6.8186296221117536e+05, + "cpu_time": 4.8153915684600395e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4604967463834816e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8174640771290229e+05, + "cpu_time": 4.8164181816491377e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4609175020521534e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8180775666994276e+05, + "cpu_time": 4.8153915684600389e+08, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.4606959712430645e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5196148328191595e+02, + "cpu_time": 2.8775016518791195e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4867410147737060e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2290030656958016e-04, + "cpu_time": 5.9743600811959909e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2295509744631119e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 561, + "real_time": 1.2809007641080760e+06, + "cpu_time": 6.1329326629411161e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.6195965323952953e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 561, + "real_time": 1.2493687169294332e+06, + "cpu_time": 6.6108685721390975e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.6857109150664940e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 561, + "real_time": 1.2691032212831865e+06, + "cpu_time": 6.2839569248840964e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.6439482177086601e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 561, + "real_time": 1.2538555605708889e+06, + "cpu_time": 6.5509607377005744e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.6761002666624886e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 561, + "real_time": 1.2758645773143135e+06, + "cpu_time": 6.1595650229411316e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.6299367971036442e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2658185680411798e+06, + "cpu_time": 6.3476567841212046e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.6510585457873165e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2691032212831865e+06, + "cpu_time": 6.2839569248840976e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.6439482177086601e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3719466432490754e+04, + "cpu_time": 2.2144727013431624e+07, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8785490271013582e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0838414587108844e-02, + "cpu_time": 3.4886459313343970e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0858111872616080e-02, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 292, + "real_time": 2.4015736576386611e+06, + "cpu_time": 8.2797429126027310e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7943704240154163e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 292, + "real_time": 2.4040263434808240e+06, + "cpu_time": 8.2335688984247041e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7915194932028954e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 292, + "real_time": 2.4015194878352117e+06, + "cpu_time": 8.2473892914040375e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7944334551494133e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 292, + "real_time": 2.4020175628159959e+06, + "cpu_time": 8.1147607866780996e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7938540100150299e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 292, + "real_time": 2.4004378013172480e+06, + "cpu_time": 8.1100649388013160e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7956926841917667e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4019149706175881e+06, + "cpu_time": 8.1971053655821776e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7939740133149044e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4015736576386611e+06, + "cpu_time": 8.2335688984247041e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7943704240154163e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3156130560672093e+03, + "cpu_time": 7.9125994733190658e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5297979930100575e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.4773506646196325e-04, + "cpu_time": 9.6529190737773239e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4753479657279700e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 147, + "real_time": 4.7766400887599206e+06, + "cpu_time": 9.0470554858502746e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8098773511496597e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 147, + "real_time": 4.7771450619967207e+06, + "cpu_time": 9.0947621436735225e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8095803300538776e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 147, + "real_time": 4.7761675565928016e+06, + "cpu_time": 9.0308392997279084e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8101553475596146e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 147, + "real_time": 4.7779253487582924e+06, + "cpu_time": 9.0577911505442262e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8091214952716053e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 147, + "real_time": 4.7768119146071728e+06, + "cpu_time": 9.0784018378911805e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8097762775538872e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7769379941429822e+06, + "cpu_time": 9.0617699835374236e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8097021603177292e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7768119146071738e+06, + "cpu_time": 9.0577911505442274e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8097762775538872e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.5497402571064549e+02, + "cpu_time": 2.5269329060598505e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8522018831416401e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3711168671515331e-04, + "cpu_time": 2.7885643871457186e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3710356697401769e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 73, + "real_time": 9.5285806192518920e+06, + "cpu_time": 9.7438937746575904e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8171609888847794e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 73, + "real_time": 9.5286618101678491e+06, + "cpu_time": 9.5882192702740252e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8171369846871655e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 73, + "real_time": 9.5278878300769683e+06, + "cpu_time": 9.6052007198630261e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8173658295243755e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 73, + "real_time": 9.5275326772299539e+06, + "cpu_time": 9.6027894280821609e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8174708509952366e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 73, + "real_time": 9.5275938893630076e+06, + "cpu_time": 9.6763672190409720e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8174527495309410e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5280513652179353e+06, + "cpu_time": 9.6432940823835540e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8173174807244995e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5278878300769683e+06, + "cpu_time": 9.6052007198630238e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8173658295243755e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3800972881198390e+02, + "cpu_time": 6.5839001873988407e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5908033935090785e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.6465872001486442e-05, + "cpu_time": 6.8274389758851807e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6465180243015726e-05, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 37, + "real_time": 1.9035296111896232e+07, + "cpu_time": 1.0330703609999994e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8203969554457260e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 37, + "real_time": 1.9036645170401882e+07, + "cpu_time": 1.0335457636216284e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8201970840677601e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 37, + "real_time": 1.9032783856665768e+07, + "cpu_time": 1.0601581630270340e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8207692371390755e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 37, + "real_time": 1.9034526587740794e+07, + "cpu_time": 1.0592469019729849e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8205109779077576e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 37, + "real_time": 1.9034223832391404e+07, + "cpu_time": 1.0629302124864852e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8205558405086231e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9034695111819219e+07, + "cpu_time": 1.0497902804216264e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8204860190137886e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9034526587740798e+07, + "cpu_time": 1.0592469019729849e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8205109779077576e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4201523144189805e+03, + "cpu_time": 1.5108098218220990e+07, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1043182481744532e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 21, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.4608618949570932e-05, + "cpu_time": 1.4391539434098339e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.4608355935416030e-05, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18, + "real_time": 3.7999729522400431e+07, + "cpu_time": 1.1301632006111205e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8256564914943428e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18, + "real_time": 3.7999906680650182e+07, + "cpu_time": 1.1285756050555449e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8256433180841385e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18, + "real_time": 3.7986065571506821e+07, + "cpu_time": 1.1298961568333476e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8266729071446899e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18, + "real_time": 3.7989780513776675e+07, + "cpu_time": 1.1297677919999988e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8263964926320557e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18, + "real_time": 3.7990643332401909e+07, + "cpu_time": 1.1279585489444597e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8263323013648846e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7993225124147207e+07, + "cpu_time": 1.1292722606888945e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8261403021440224e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7990643332401916e+07, + "cpu_time": 1.1297677919999986e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8263323013648846e+10, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2598038397934524e+03, + "cpu_time": 9.5390503228971292e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6562625993300676e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 22, + "run_name": "Comm_hipMemcpyAsync_PinnedToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6476105461799644e-04, + "cpu_time": 8.4470775161677879e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6475695123124786e-04, + "hip_id": NaN, + "numa_id": NaN + } + ] +} diff --git a/results/implicit_managed_GPURdHost_coarse.json b/results/implicit_managed_GPURdHost_coarse.json new file mode 100644 index 0000000..49cd4c4 --- /dev/null +++ b/results/implicit_managed_GPURdHost_coarse.json @@ -0,0 +1,3194 @@ +{ + "context": { + "date": "2023-10-09T18:18:41-04:00", + "host_name": "frontier08444", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [2.44,2.61,4.37], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8648067624968779e+06, + "cpu_time": 2.0648406243386250e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1964742312043486e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8610708470097200e+06, + "cpu_time": 2.0649065132275133e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.2008834357817476e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8693731853904119e+06, + "cpu_time": 2.0691600449735466e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1911087802110338e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8649951499937933e+06, + "cpu_time": 2.0646817010582001e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1962523602346266e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8606357440129465e+06, + "cpu_time": 2.0647274232804223e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.2013981044811639e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8641763377807499e+06, + "cpu_time": 2.0656632613756615e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1972233823825843e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8648067624968779e+06, + "cpu_time": 2.0648406243386250e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1964742312043486e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5445377342974380e+03, + "cpu_time": 1.9567925655588140e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1748283407590170e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9013961621876993e-03, + "cpu_time": 9.4729504181415155e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9000472934308547e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8644300473853331e+06, + "cpu_time": 2.0652232222222232e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3938360741870785e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8596973342321464e+06, + "cpu_time": 2.0654901507936483e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.4050178753320677e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8637756724726709e+06, + "cpu_time": 2.0651562407407425e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3953787577512879e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8601180756820005e+06, + "cpu_time": 2.0655041957671945e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.4040215011600563e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8597879068924128e+06, + "cpu_time": 2.0653520820105844e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.4048033486185586e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8615618073329129e+06, + "cpu_time": 2.0653451783068788e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.4006115114098107e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8601180756820005e+06, + "cpu_time": 2.0653520820105844e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.4040215011600563e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3364151840360987e+03, + "cpu_time": 1.5566143114844516e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5205034343749221e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2550833256423086e-03, + "cpu_time": 7.5368240032425347e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2544855232190981e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8661257546761865e+06, + "cpu_time": 2.0682437222222215e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7796869846228454e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8607013398121116e+06, + "cpu_time": 2.0675510740740723e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.8052819920333959e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8644182408352015e+06, + "cpu_time": 2.0674443941798911e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7877277968812808e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8602370170892708e+06, + "cpu_time": 2.0676698439153426e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.8074798262192365e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8649194043107510e+06, + "cpu_time": 2.0673281957671980e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7853662534308322e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8632803513447042e+06, + "cpu_time": 2.0676474460317451e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7931085706375185e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8644182408352015e+06, + "cpu_time": 2.0675510740740721e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7877277968812808e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6453161960311218e+03, + "cpu_time": 3.5654077604780406e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2487597332891444e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4197091672876992e-03, + "cpu_time": 1.7243789637932789e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4201573007514984e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 373, + "real_time": 1.8673966829986472e+06, + "cpu_time": 2.0710252091152796e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7547423264874533e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 373, + "real_time": 1.8629985489489587e+06, + "cpu_time": 2.0714415656836429e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7588848911603611e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 373, + "real_time": 1.8668703551473778e+06, + "cpu_time": 2.0712717050938394e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7552370420180127e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 373, + "real_time": 1.8629668453583762e+06, + "cpu_time": 2.0713296139410157e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7589148235054322e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 373, + "real_time": 1.8669835951671770e+06, + "cpu_time": 2.0712617372654155e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7551305798734579e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8654432055241074e+06, + "cpu_time": 2.0712659662198387e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7565819326089438e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8668703551473778e+06, + "cpu_time": 2.0712717050938394e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7552370420180127e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2546804640957052e+03, + "cpu_time": 1.5239833262983336e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1239882392859428e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2086567189067753e-03, + "cpu_time": 7.3577384611773328e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2091597891658334e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 374, + "real_time": 1.8696007620922059e+06, + "cpu_time": 2.0773643101604285e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5053473088372581e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 374, + "real_time": 1.8651429145490467e+06, + "cpu_time": 2.0778865427807472e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5137253820490889e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 374, + "real_time": 1.8698107834841886e+06, + "cpu_time": 2.0776337700534768e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5049535802698068e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 374, + "real_time": 1.8655955613465335e+06, + "cpu_time": 2.0781554732620346e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5128728518574513e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 374, + "real_time": 1.8700520591022831e+06, + "cpu_time": 2.0779550160427734e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5045013683448203e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8680404161148514e+06, + "cpu_time": 2.0777990224598921e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5082800982716851e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8696007620922059e+06, + "cpu_time": 2.0778865427807469e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5053473088372581e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4488986229735824e+03, + "cpu_time": 3.0632259272775588e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6013660108303709e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3109452032450128e-03, + "cpu_time": 1.4742647841132521e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3115731589097412e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 377, + "real_time": 1.8790329355899051e+06, + "cpu_time": 2.0951603368700265e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9755030642318755e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 377, + "real_time": 1.8700660362395651e+06, + "cpu_time": 2.0905375809018551e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0089503504147366e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 377, + "real_time": 1.8762738437889593e+06, + "cpu_time": 2.0935442015915138e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9857606571603835e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 377, + "real_time": 1.8711393336294247e+06, + "cpu_time": 2.0881779602122032e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0049299720380172e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 377, + "real_time": 1.8759181648347680e+06, + "cpu_time": 2.0929391962864741e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9870851755169660e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8744860628165249e+06, + "cpu_time": 2.0920718551724148e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9924458438723966e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8759181648347680e+06, + "cpu_time": 2.0929391962864741e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9870851755169660e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7635873911800290e+03, + "cpu_time": 2.7374814741047749e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4042243951733640e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0077969454331498e-03, + "cpu_time": 1.3085026058434160e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0082020319169302e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 376, + "real_time": 1.8707361730853433e+06, + "cpu_time": 2.1079484680851074e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.4012879195448205e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 376, + "real_time": 1.8775200830745273e+06, + "cpu_time": 2.1113341223404217e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3962247454137847e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 376, + "real_time": 1.8767804861628509e+06, + "cpu_time": 2.1109846648936118e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3967749661334309e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 376, + "real_time": 1.8779481467410075e+06, + "cpu_time": 2.1113266888297824e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3959064868480256e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 376, + "real_time": 1.8802115801589296e+06, + "cpu_time": 2.1133979148936248e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3942260688440266e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8766392938445318e+06, + "cpu_time": 2.1109983718085098e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3968840373568174e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8775200830745273e+06, + "cpu_time": 2.1113266888297824e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3962247454137847e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5396640042457839e+03, + "cpu_time": 1.9545094082364278e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6391827494510490e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8861717410778162e-03, + "cpu_time": 9.2586969006611954e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8893356061573357e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 372, + "real_time": 1.8965987688442443e+06, + "cpu_time": 2.1628780591397770e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7643590653572577e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 372, + "real_time": 1.8908004268630838e+06, + "cpu_time": 2.1575007983870828e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7728362684464562e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 372, + "real_time": 1.8904735761386673e+06, + "cpu_time": 2.1574215107526937e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7733156740062428e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 372, + "real_time": 1.8929359388865673e+06, + "cpu_time": 2.1602722338709603e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7697080985655987e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 372, + "real_time": 1.8956099455501453e+06, + "cpu_time": 2.1621238225806425e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7658010617149442e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8932837312565416e+06, + "cpu_time": 2.1600392849462312e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7692040336181003e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8929359388865673e+06, + "cpu_time": 2.1602722338709603e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7697080985655987e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7652070074181111e+03, + "cpu_time": 2.5374621677357186e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0437214671636326e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4605349223504329e-03, + "cpu_time": 1.1747296382152985e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4602468500236557e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 368, + "real_time": 1.9129451950151555e+06, + "cpu_time": 2.2381212608695687e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4814743398421955e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 368, + "real_time": 1.9080894975480116e+06, + "cpu_time": 2.2350753315217514e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4954235707888520e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 368, + "real_time": 1.9133664918626889e+06, + "cpu_time": 2.2379742201087065e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4802673949787664e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 368, + "real_time": 1.9068620646145933e+06, + "cpu_time": 2.2342474755434897e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4989609340827370e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 368, + "real_time": 1.9144548041367671e+06, + "cpu_time": 2.2388767500000121e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4771520212137151e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9111436106354434e+06, + "cpu_time": 2.2368590076087057e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4866556521812534e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9129451950151552e+06, + "cpu_time": 2.2379742201087065e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4814743398421955e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4208972766209326e+03, + "cpu_time": 2.0560804143502164e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.8268915304891812e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7899739494111097e-03, + "cpu_time": 9.1918194546747550e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7910530846933033e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 358, + "real_time": 1.9595538810358574e+06, + "cpu_time": 2.4178605027932944e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0702191046114055e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 358, + "real_time": 1.9586479721211062e+06, + "cpu_time": 2.4180118743016850e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0707140996495156e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 358, + "real_time": 1.9581921079310498e+06, + "cpu_time": 2.4169702849161969e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0709633602883681e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 358, + "real_time": 1.9597590383267102e+06, + "cpu_time": 2.4184585502793407e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0701070687703520e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 358, + "real_time": 1.9560512782896571e+06, + "cpu_time": 2.4133920949720680e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0721354921910428e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9584408555408760e+06, + "cpu_time": 2.4169386614525169e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0708278251021369e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9586479721211065e+06, + "cpu_time": 2.4178605027932944e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0707140996495156e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4825049492010874e+03, + "cpu_time": 2.0549192909473395e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1106824719008699e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5698224176989697e-04, + "cpu_time": 8.5021573932388779e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5742171446910832e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 336, + "real_time": 2.0886270867777055e+06, + "cpu_time": 2.7940279970238050e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0081631740546336e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 336, + "real_time": 2.0956166283992517e+06, + "cpu_time": 2.8003795238095312e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0014653172531095e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 336, + "real_time": 2.0893180165633871e+06, + "cpu_time": 2.7943975624999949e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0074990818769646e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 336, + "real_time": 2.0890575258471514e+06, + "cpu_time": 2.7932564940476129e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0077494028313711e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 336, + "real_time": 2.0892365670685347e+06, + "cpu_time": 2.7945561309523741e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0075773448121979e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0903711649312060e+06, + "cpu_time": 2.7953235416666637e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0064908641656556e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0892365670685344e+06, + "cpu_time": 2.7943975624999953e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0075773448121979e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9444344550504084e+03, + "cpu_time": 2.8705589101259984e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8210638441552506e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4085701642116313e-03, + "cpu_time": 1.0269147264486160e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4059689453549110e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 314, + "real_time": 2.2262953485771539e+06, + "cpu_time": 3.4548329394904450e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7679672669492111e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 314, + "real_time": 2.2315106860356061e+06, + "cpu_time": 3.4576669777069935e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7591610259786811e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 314, + "real_time": 2.2233449883616655e+06, + "cpu_time": 3.4526648312101783e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7729673280174947e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 314, + "real_time": 2.2301369407481165e+06, + "cpu_time": 3.4567410382165671e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7614766370292840e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 314, + "real_time": 2.2248323816972175e+06, + "cpu_time": 3.4547048025477692e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7704449418345551e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2272240690839519e+06, + "cpu_time": 3.4553221178343915e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7664034399618454e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2262953485771539e+06, + "cpu_time": 3.4548329394904450e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7679672669492111e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4817331819845081e+03, + "cpu_time": 1.9489259017035881e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8862098602615250e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5632612947724341e-03, + "cpu_time": 5.6403595243533161e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5628197972124712e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 210, + "real_time": 3.3295433059157361e+06, + "cpu_time": 5.6089175809523864e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0388940639970751e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 210, + "real_time": 3.3301346103793806e+06, + "cpu_time": 5.6104858523809751e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0379993492481318e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 210, + "real_time": 3.3388836131918998e+06, + "cpu_time": 5.6174360714285923e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0247980893114595e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 210, + "real_time": 3.3304111100733280e+06, + "cpu_time": 5.6087122380952379e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0375810809827032e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 210, + "real_time": 3.3314077083819681e+06, + "cpu_time": 5.6085922047619047e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0360740769698610e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3320760695884628e+06, + "cpu_time": 5.6108287895238195e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0350693321018467e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3304111100733280e+06, + "cpu_time": 5.6089175809523864e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0375810809827032e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8647639797008828e+03, + "cpu_time": 3.7718894345740910e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8316035034491159e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1598666714047170e-03, + "cpu_time": 6.7225174320355698e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1581972598210008e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 104, + "real_time": 6.6526710243824013e+06, + "cpu_time": 1.0952665192307701e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0437533852224445e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 104, + "real_time": 6.6722296402216535e+06, + "cpu_time": 1.0970076971153922e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0289683972695694e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 104, + "real_time": 6.6576325200283183e+06, + "cpu_time": 1.0960438711538408e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0399946075511656e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 104, + "real_time": 6.6694835012850277e+06, + "cpu_time": 1.0967630932692228e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0310390592517357e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 104, + "real_time": 6.6715926558782272e+06, + "cpu_time": 1.0977876605769219e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0294485486064196e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6647218683591262e+06, + "cpu_time": 1.0965737682692297e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0346407995802670e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6694835012850286e+06, + "cpu_time": 1.0967630932692230e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0310390592517357e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9684057659686623e+03, + "cpu_time": 9.6012547801365145e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7788389201927492e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3456534185689450e-03, + "cpu_time": 8.7556852607285993e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3464394363065374e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52, + "real_time": 1.3572544647523990e+07, + "cpu_time": 2.2267987884615462e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9444570449243298e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52, + "real_time": 1.3253713814684978e+07, + "cpu_time": 2.1369407249999847e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0634007145713434e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52, + "real_time": 1.3583027883074604e+07, + "cpu_time": 2.2290338423076969e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9406409658940849e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52, + "real_time": 1.3578680039455112e+07, + "cpu_time": 2.2278848615384616e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9422229410372763e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52, + "real_time": 1.3542058143334894e+07, + "cpu_time": 2.2242661134615269e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9555882340550671e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3506004905614719e+07, + "cpu_time": 2.2089848661538433e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9692619800964203e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3572544647523990e+07, + "cpu_time": 2.2267987884615462e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9444570449243298e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4194398710524556e+05, + "cpu_time": 4.0312472921602387e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2949574019689336e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0509694620815413e-02, + "cpu_time": 1.8249320554102361e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0655420107003885e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27, + "real_time": 2.6283959961599775e+07, + "cpu_time": 4.2580222370370246e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1064500248854752e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27, + "real_time": 2.6292765857996766e+07, + "cpu_time": 4.2623691592592858e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1047397875480108e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27, + "real_time": 2.6265470380032506e+07, + "cpu_time": 4.2567152296295829e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1100447110985222e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27, + "real_time": 2.6286946609616280e+07, + "cpu_time": 4.2653314962963045e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1058698445752745e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27, + "real_time": 2.6110613235720880e+07, + "cpu_time": 4.1675103074074067e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1403514267670326e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6247951208993241e+07, + "cpu_time": 4.2419896859259203e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1134911589748640e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6283959961599775e+07, + "cpu_time": 4.2580222370370239e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1064500248854752e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7450410836494513e+04, + "cpu_time": 4.1776863219146163e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5145960625666501e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9507221428374934e-03, + "cpu_time": 9.8484122575199801e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9619608511659017e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10, + "real_time": 5.1235353946685791e+07, + "cpu_time": 8.1178326299999520e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.2392622539375277e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10, + "real_time": 5.1293835416436195e+07, + "cpu_time": 8.1084651799999103e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.2332888313121662e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10, + "real_time": 4.7543637081980705e+07, + "cpu_time": 7.9450426600000411e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.6460858376722403e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10, + "real_time": 5.1264522597193718e+07, + "cpu_time": 8.1066110899999440e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.2362812018987665e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10, + "real_time": 5.1219387352466583e+07, + "cpu_time": 8.1066656000000134e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.2408954865617485e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0511347278952599e+07, + "cpu_time": 8.0769234319999725e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3191627222764902e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1235353946685791e+07, + "cpu_time": 8.1066656000000134e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.2392622539375277e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6592446292773630e+06, + "cpu_time": 7.3869709566962335e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8277871865779036e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.2848948180179464e-02, + "cpu_time": 9.1457731633678551e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4362310047842438e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 1.0126967169344425e+08, + "cpu_time": 1.5925211775000036e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.3013987605803070e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.9620835147798061e+07, + "cpu_time": 1.5567606137500077e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.9904698624446001e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 1.0153957735747099e+08, + "cpu_time": 1.5948143000000137e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.2873069395388670e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.9733739383518696e+07, + "cpu_time": 1.5653525349999952e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.9829325701610794e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 1.0150537919253111e+08, + "cpu_time": 1.5931481287499949e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.2890882854758463e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6733840554952621e+07, + "cpu_time": 1.5805193510000032e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.5702392836401405e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0126967169344425e+08, + "cpu_time": 1.5925211775000036e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.3013987605803070e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4426840997056458e+06, + "cpu_time": 1.8044288133633323e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8022404909745514e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.6602174200306688e-02, + "cpu_time": 1.1416682827841749e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8259913037161210e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.8767516687512398e+08, + "cpu_time": 2.8531457724999851e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7212781098228645e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.8765720352530479e+08, + "cpu_time": 2.8563844199999976e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7218257750239277e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.8773095682263374e+08, + "cpu_time": 2.8582995374999952e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7195778585119562e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.8751276284456253e+08, + "cpu_time": 2.8545517099999887e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7262332851981449e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.8726067990064621e+08, + "cpu_time": 2.8510142275000304e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7339417146711674e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8756735399365425e+08, + "cpu_time": 2.8546791334999996e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7245713486456118e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8765720352530479e+08, + "cpu_time": 2.8545517099999887e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7218257750239277e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8940144816571317e+05, + "cpu_time": 2.8202081426441885e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7854198794970792e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0097783230024185e-03, + "cpu_time": 9.8792474066479495e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0106293601993210e-03, + "dst_gpu": NaN, + "src_numa": NaN + } + ] +} diff --git a/results/implicit_managed_GPURdHost_fine.json b/results/implicit_managed_GPURdHost_fine.json new file mode 100644 index 0000000..19fa59f --- /dev/null +++ b/results/implicit_managed_GPURdHost_fine.json @@ -0,0 +1,3194 @@ +{ + "context": { + "date": "2023-10-09T18:16:28-04:00", + "host_name": "frontier08444", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1883, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [2.1,2.56,4.64], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8639331348591468e+06, + "cpu_time": 2.0661555740740735e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1975037212424069e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8663547286388364e+06, + "cpu_time": 2.0662311931216945e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1946524619075395e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8621231467198525e+06, + "cpu_time": 2.0661244867724872e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1996397000999334e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8622628324404890e+06, + "cpu_time": 2.0661392195767192e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1994747082140958e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8657714362710803e+06, + "cpu_time": 2.0651747433862439e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1953385716882027e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8640890557858809e+06, + "cpu_time": 2.0659650433862437e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1973218326304359e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8639331348591466e+06, + "cpu_time": 2.0661392195767190e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1975037212424069e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9485814756939285e+03, + "cpu_time": 4.4369788478597695e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2967003706999826e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0453263859072583e-03, + "cpu_time": 2.1476543671752004e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0452271199392670e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8595415241617484e+06, + "cpu_time": 2.0650127724867745e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.4053869696149016e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8642683220768862e+06, + "cpu_time": 2.0661589682539690e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3942172395407706e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8590687233506560e+06, + "cpu_time": 2.0648353227513230e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.4065073534427006e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8639009264661197e+06, + "cpu_time": 2.0648539021164030e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3950833886496844e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8680520668413637e+06, + "cpu_time": 2.0689611534391560e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3853167400476271e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8629663125793547e+06, + "cpu_time": 2.0659644238095251e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3973023382591372e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8639009264661192e+06, + "cpu_time": 2.0650127724867743e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3950833886496844e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7200136344653997e+03, + "cpu_time": 1.7629523450286561e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7776663275170449e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9968228138891497e-03, + "cpu_time": 8.5333141496109048e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9961480135550708e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8606246864437321e+06, + "cpu_time": 2.0678239682539701e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.8056447489768770e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8668474032276890e+06, + "cpu_time": 2.0690159126984111e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7762931087312512e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8623965543012011e+06, + "cpu_time": 2.0692205820105830e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7972671352731977e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8643292562777998e+06, + "cpu_time": 2.0672384259259254e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7881472357040849e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8706721915920740e+06, + "cpu_time": 2.0723980740740763e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7583490435360894e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8649740183684994e+06, + "cpu_time": 2.0691393925925933e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7851402544442993e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8643292562778001e+06, + "cpu_time": 2.0690159126984111e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7881472357040849e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9355582920648076e+03, + "cpu_time": 1.9990534925075845e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8524162870128908e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1102483215866353e-03, + "cpu_time": 9.6612799488719195e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1085790702951783e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 377, + "real_time": 1.8627436413602293e+06, + "cpu_time": 2.0712722838196279e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7591255861741584e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 377, + "real_time": 1.8669720251424748e+06, + "cpu_time": 2.0708424058355396e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7551414567928173e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 377, + "real_time": 1.8631930971406687e+06, + "cpu_time": 2.0713325809018505e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7587012344714619e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 377, + "real_time": 1.8685448763881866e+06, + "cpu_time": 2.0726020450928318e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7536640630937949e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 377, + "real_time": 1.8721727171422734e+06, + "cpu_time": 2.0770472466843554e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7502658648940153e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8667252714347667e+06, + "cpu_time": 2.0726193124668412e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7553796410852492e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8669720251424748e+06, + "cpu_time": 2.0713325809018507e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7551414567928173e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9171465254389600e+03, + "cpu_time": 2.5609793570062939e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6815667105909204e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0984054726104597e-03, + "cpu_time": 1.2356245749530358e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0973051210248866e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 377, + "real_time": 1.8640516733979825e+06, + "cpu_time": 2.0778099496021236e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5157823645808235e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 377, + "real_time": 1.8686522576340046e+06, + "cpu_time": 2.0775328514588883e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5071265791837834e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 377, + "real_time": 1.8666142402327959e+06, + "cpu_time": 2.0798612493368678e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5109557501193523e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 377, + "real_time": 1.8681166500379012e+06, + "cpu_time": 2.0770835198938998e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5081321072038181e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 377, + "real_time": 1.8745201208831901e+06, + "cpu_time": 2.0830701352785148e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.4961481218522407e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8683909884371746e+06, + "cpu_time": 2.0790715411140588e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5076289845880032e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8681166500379010e+06, + "cpu_time": 2.0778099496021238e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5081321072038181e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8631682879117316e+03, + "cpu_time": 2.4760756116300536e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.2432852862150889e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0676444661848312e-03, + "cpu_time": 1.1909525779489351e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0650089613357056e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 377, + "real_time": 1.8761056749087165e+06, + "cpu_time": 2.0931567639257340e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9863868412624165e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 377, + "real_time": 1.8720449762423807e+06, + "cpu_time": 2.0883560397877956e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0015411842877448e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 377, + "real_time": 1.8753298874885777e+06, + "cpu_time": 2.0928386074270559e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9892769733185589e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 377, + "real_time": 1.8689060667934222e+06, + "cpu_time": 2.0897950769230728e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0133005788186535e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 377, + "real_time": 1.8781483877729378e+06, + "cpu_time": 2.0955117135278527e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9787883030595869e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8741069986412073e+06, + "cpu_time": 2.0919316403183024e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9938587761493921e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8753298874885780e+06, + "cpu_time": 2.0928386074270562e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9892769733185589e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6443318498488106e+03, + "cpu_time": 2.8500115643662684e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3610152214431696e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9445697884331466e-03, + "cpu_time": 1.3623827420730719e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9460147323599570e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 376, + "real_time": 1.8809357711592035e+06, + "cpu_time": 2.1140755425531878e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3936892690303990e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 376, + "real_time": 1.8813663912024298e+06, + "cpu_time": 2.1145230824468145e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3933702718716952e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 376, + "real_time": 1.8800115210863021e+06, + "cpu_time": 2.1139059122340446e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3943744336658573e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 376, + "real_time": 1.8811987487656842e+06, + "cpu_time": 2.1138088723404203e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3934944416266555e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 376, + "real_time": 1.8797825753395916e+06, + "cpu_time": 2.1127366728723445e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3945442597404778e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8806590015106425e+06, + "cpu_time": 2.1138100164893628e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3938945351870170e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8809357711592033e+06, + "cpu_time": 2.1139059122340446e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3936892690303990e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1688256523339908e+02, + "cpu_time": 6.5958907126286692e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3139100952123757e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.8118689494350753e-04, + "cpu_time": 3.1203801009435997e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8122755782950361e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 372, + "real_time": 1.8934863956703963e+06, + "cpu_time": 2.1593213333333330e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7689029147440678e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 372, + "real_time": 1.8991144349446809e+06, + "cpu_time": 2.1658830188172143e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7606972510599232e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 372, + "real_time": 1.8947706849630971e+06, + "cpu_time": 2.1618446962365638e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7670261322953236e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 372, + "real_time": 1.8984276141191493e+06, + "cpu_time": 2.1622007903225860e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7616960272844756e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 372, + "real_time": 1.8974244776864096e+06, + "cpu_time": 2.1621266209677514e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7631560895603138e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8966447214767467e+06, + "cpu_time": 2.1622752919354895e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7642956829888207e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8974244776864096e+06, + "cpu_time": 2.1621266209677509e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7631560895603138e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4172942149537644e+03, + "cpu_time": 2.3427884494987370e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5244692215991899e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2745108177516951e-03, + "cpu_time": 1.0834829673337602e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2749971876338684e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 370, + "real_time": 1.9156653540227814e+06, + "cpu_time": 2.2391732081081159e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4736908917732096e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 370, + "real_time": 1.9078632877239755e+06, + "cpu_time": 2.2350291918918877e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4960751472445393e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 370, + "real_time": 1.9143507153588675e+06, + "cpu_time": 2.2389469999999860e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4774498297895861e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 370, + "real_time": 1.9083886993325292e+06, + "cpu_time": 2.2351333081081053e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4945619850229990e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 370, + "real_time": 1.9130512788841450e+06, + "cpu_time": 2.2367939054054073e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4811703772604525e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9118638670644599e+06, + "cpu_time": 2.2370153227027012e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4845896462181568e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9130512788841452e+06, + "cpu_time": 2.2367939054054073e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4811703772604525e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5400243710578993e+03, + "cpu_time": 1.9952172031223658e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0158930829176857e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8516090146592770e-03, + "cpu_time": 8.9191038741379700e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8522681703601737e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 360, + "real_time": 1.9570612578213010e+06, + "cpu_time": 2.4160916166666662e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0715821958146856e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 360, + "real_time": 1.9570532364822510e+06, + "cpu_time": 2.4159923750000088e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0715865878894397e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 360, + "real_time": 1.9576061465260056e+06, + "cpu_time": 2.4160133888888899e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0712839269133040e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 360, + "real_time": 1.9588537155262504e+06, + "cpu_time": 2.4165216722222185e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0706016398149441e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 360, + "real_time": 1.9557621431886218e+06, + "cpu_time": 2.4114770194444386e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0722939940850168e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9572672999088862e+06, + "cpu_time": 2.4152192144444445e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0714696689034781e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9570612578213005e+06, + "cpu_time": 2.4160133888888895e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0715821958146856e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1163898043329150e+03, + "cpu_time": 2.1029738235612122e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1110943059324485e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.7038188109763280e-04, + "cpu_time": 8.7071757751187992e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7034692472316339e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 336, + "real_time": 2.0902584098318282e+06, + "cpu_time": 2.7952548511904655e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0065959214762604e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 336, + "real_time": 2.0957070300937628e+06, + "cpu_time": 2.7998476011904790e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0013789808265064e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 336, + "real_time": 2.0903541416850029e+06, + "cpu_time": 2.7944873392857043e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0065040254944718e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 336, + "real_time": 2.0911607688980266e+06, + "cpu_time": 2.7954600297618895e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0057300530796881e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 336, + "real_time": 2.0888064717014674e+06, + "cpu_time": 2.7939864642857099e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0079907147087059e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0912573644420176e+06, + "cpu_time": 2.7958072571428502e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0056399391171265e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0903541416850027e+06, + "cpu_time": 2.7952548511904655e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0065040254944718e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6281862566948698e+03, + "cpu_time": 2.3349287453291495e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5174679491266618e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2567493132994245e-03, + "cpu_time": 8.3515368928375588e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2551943646648958e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 314, + "real_time": 2.2257530975013166e+06, + "cpu_time": 3.4542125732484106e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7688852413222523e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 314, + "real_time": 2.2328818223593389e+06, + "cpu_time": 3.4582310636942554e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7568526538212900e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 314, + "real_time": 2.2267446802057042e+06, + "cpu_time": 3.4552940891719661e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7672069342162166e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 314, + "real_time": 2.2312375096972012e+06, + "cpu_time": 3.4567987898089052e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7596212700540380e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 314, + "real_time": 2.2255003972833226e+06, + "cpu_time": 3.4543995796178328e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7693131891775923e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2284235014093770e+06, + "cpu_time": 3.4557872191082737e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7643758577182779e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2267446802057042e+06, + "cpu_time": 3.4552940891719661e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.7672069342162166e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4018121543865964e+03, + "cpu_time": 1.7061566251051941e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7430714190927492e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5265555008889039e-03, + "cpu_time": 4.9370997602839911e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5256370873055777e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 210, + "real_time": 3.3311333329904648e+06, + "cpu_time": 5.6086583285714183e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0364888831809559e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 210, + "real_time": 3.3536958264275678e+06, + "cpu_time": 5.6312421952381236e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0026051461773348e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 210, + "real_time": 3.3282403645682195e+06, + "cpu_time": 5.6034556714285780e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0408666929849424e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 210, + "real_time": 3.3311637938909587e+06, + "cpu_time": 5.6077691333333254e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0364428284096498e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 210, + "real_time": 3.3290906665137126e+06, + "cpu_time": 5.6042071761904741e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0395791766072330e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3346647968781847e+06, + "cpu_time": 5.6110665009523835e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0311965454720230e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3311333329904643e+06, + "cpu_time": 5.6077691333333245e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.0364888831809559e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0715089881027667e+04, + "cpu_time": 1.1496799876121886e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6099539546091372e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.2132434693462500e-03, + "cpu_time": 2.0489509212144427e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1999424790074317e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 104, + "real_time": 6.6755048867959818e+06, + "cpu_time": 1.0974598624999933e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0265010016500788e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 104, + "real_time": 6.6476185725500379e+06, + "cpu_time": 1.0940432644230733e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0475868363681498e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 104, + "real_time": 6.6554201587748071e+06, + "cpu_time": 1.0963515057692366e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0416699771779728e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 104, + "real_time": 6.6440938837611331e+06, + "cpu_time": 1.0944228692307634e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0502645788932295e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 104, + "real_time": 6.6343799430447128e+06, + "cpu_time": 1.0912445846153755e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0576590861633530e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6514034889853355e+06, + "cpu_time": 1.0947044173076885e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0447362960505571e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6476185725500379e+06, + "cpu_time": 1.0944228692307634e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0475868363681498e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5448042351495029e+04, + "cpu_time": 2.3870595574563948e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1700579781436473e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3225237165474697e-03, + "cpu_time": 2.1805516810895122e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3193640053290137e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52, + "real_time": 1.3561138166831089e+07, + "cpu_time": 2.2249436942307565e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9486159033568583e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52, + "real_time": 1.3205759101905502e+07, + "cpu_time": 2.1306824538461421e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0817876868825092e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52, + "real_time": 1.3570153691734258e+07, + "cpu_time": 2.2263162480768982e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9453282198916292e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52, + "real_time": 1.3581969130497713e+07, + "cpu_time": 2.2273357346153833e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9410261027106895e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52, + "real_time": 1.3564654984153233e+07, + "cpu_time": 2.2246172519230682e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9473329088280706e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3496735015024360e+07, + "cpu_time": 2.2067790765384499e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9728181643339520e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3564654984153235e+07, + "cpu_time": 2.2249436942307565e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.9473329088280706e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6285194711345673e+05, + "cpu_time": 4.2553281557225919e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0983679695320450e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2066025370741321e-02, + "cpu_time": 1.9282982156951994e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2263404307180909e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27, + "real_time": 2.6271017406273771e+07, + "cpu_time": 4.2577794851851530e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1089657444308767e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27, + "real_time": 2.6279574467076197e+07, + "cpu_time": 4.2616520740740731e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1073021813253412e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27, + "real_time": 2.6268800927533045e+07, + "cpu_time": 4.2588406962963171e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1093968228798275e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27, + "real_time": 2.6288160916279864e+07, + "cpu_time": 4.2659885814815000e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1056339934712200e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27, + "real_time": 2.6116420411401324e+07, + "cpu_time": 4.1649207074074097e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1392084323089781e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6244794825712841e+07, + "cpu_time": 4.2418363088888906e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1141014348832493e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6271017406273775e+07, + "cpu_time": 4.2588406962963179e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.1089657444308767e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2170194971769000e+04, + "cpu_time": 4.3114210633046424e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4113800190304523e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7498860422052762e-03, + "cpu_time": 1.0164043940757296e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7597810426743578e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10, + "real_time": 5.1197369769215584e+07, + "cpu_time": 8.1089192299999982e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.2431493494692631e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10, + "real_time": 5.1283114030957222e+07, + "cpu_time": 8.1127683399999738e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.2343829167229986e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10, + "real_time": 4.7494085133075714e+07, + "cpu_time": 7.9441692200001061e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.6519765618783722e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10, + "real_time": 5.1250937953591347e+07, + "cpu_time": 8.1082567900000423e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.2376691377448187e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10, + "real_time": 5.1259352639317513e+07, + "cpu_time": 8.1083335300000668e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.2368093270476007e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0496971905231483e+07, + "cpu_time": 8.0764894220000371e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3207974585726109e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1250937953591347e+07, + "cpu_time": 8.1083335300000668e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.2376691377448187e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6789576202130925e+06, + "cpu_time": 7.3992737953478657e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8516251261901152e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.3248679215142254e-02, + "cpu_time": 9.1614975377699824e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4799767151573617e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 1.0131103266030550e+08, + "cpu_time": 1.5957245337500048e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.2992344259299059e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.9731236919760704e+07, + "cpu_time": 1.5570127250000000e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.9830994247864838e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 1.0145877487957478e+08, + "cpu_time": 1.5933221912499994e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.2915177877638702e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.9740275405347347e+07, + "cpu_time": 1.5650059737500045e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.9824968173432808e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 1.0140473488718271e+08, + "cpu_time": 1.5930659237499967e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.2943377111265345e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6729210950434208e+07, + "cpu_time": 1.5808262695000011e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.5701372333900146e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0131103266030550e+08, + "cpu_time": 1.5930659237499967e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.2992344259299059e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3843411917624380e+06, + "cpu_time": 1.8339043960048649e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7671630607550102e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.6002204804853512e-02, + "cpu_time": 1.1600923082995767e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7631422762313070e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.8722555041313171e+08, + "cpu_time": 2.8493250325000119e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7350175851035414e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.8732339516282082e+08, + "cpu_time": 2.8548624900000161e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7320220096732044e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.8732090666890144e+08, + "cpu_time": 2.8541410375000000e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7320981576172352e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.8725286796689034e+08, + "cpu_time": 2.8544667499999934e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7341809268836231e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.8744419142603874e+08, + "cpu_time": 2.8576401525000250e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7283280737118721e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8731338232755661e+08, + "cpu_time": 2.8540870925000089e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7323293505978956e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8732090666890144e+08, + "cpu_time": 2.8544667499999934e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7320981576172352e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.4623173347457938e+04, + "cpu_time": 3.0020718663290853e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5890666690836679e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPURdHost_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.5177323849438915e-04, + "cpu_time": 1.0518501254632177e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5166048751431597e-04, + "dst_gpu": NaN, + "src_numa": NaN + } + ] +} diff --git a/results/implicit_managed_GPUWrGPU_coarse.json b/results/implicit_managed_GPUWrGPU_coarse.json new file mode 100644 index 0000000..52b42f0 --- /dev/null +++ b/results/implicit_managed_GPUWrGPU_coarse.json @@ -0,0 +1,25272 @@ +{ + "context": { + "date": "2023-10-09T17:57:28-04:00", + "host_name": "frontier08444", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [2.09,2.66,9.64], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52604, + "real_time": 1.3319161775119794e+04, + "cpu_time": 2.5218342863660549e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0752686011002070e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52604, + "real_time": 1.3305676242048889e+04, + "cpu_time": 2.5182800395407201e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0783854390322018e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52604, + "real_time": 1.3306235573199010e+04, + "cpu_time": 2.5293309672268264e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0782560382818043e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52604, + "real_time": 1.3313799540920885e+04, + "cpu_time": 2.5472604307657235e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0765071889588398e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52604, + "real_time": 1.3306945326532652e+04, + "cpu_time": 2.5362465287810821e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0780918531565660e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3310363691564247e+04, + "cpu_time": 2.5305904505360813e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0773018241059238e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3306945326532650e+04, + "cpu_time": 2.5293309672268264e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0780918531565660e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.9141799593223716e+00, + "cpu_time": 1.1615532596723934e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3669612927950814e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.4432895271453928e-04, + "cpu_time": 4.5900483795247443e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4420774136844278e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52591, + "real_time": 1.3303327701800850e+04, + "cpu_time": 2.5685824209465496e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1578577808701670e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52591, + "real_time": 1.3308652240537440e+04, + "cpu_time": 2.5694487592934151e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1553941390455806e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52591, + "real_time": 1.3307618157962650e+04, + "cpu_time": 2.5765962693236481e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1558724504717577e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52591, + "real_time": 1.3304267059251250e+04, + "cpu_time": 2.5609796790325330e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1574230008436394e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52591, + "real_time": 1.3302335702590970e+04, + "cpu_time": 2.5635055712954676e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1583169927100837e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3305240172428636e+04, + "cpu_time": 2.5678225399783223e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1569728727882469e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3304267059251250e+04, + "cpu_time": 2.5685824209465496e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1574230008436394e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7539726785078615e+00, + "cpu_time": 6.0345111547607019e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2743297657984766e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0698406363341680e-04, + "cpu_time": 2.3500499200430121e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0697342543615652e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52610, + "real_time": 1.3311058479971016e+04, + "cpu_time": 2.6004237901539604e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2308562857456303e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52610, + "real_time": 1.3306710123221137e+04, + "cpu_time": 2.5821369834632205e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2312585040391598e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52610, + "real_time": 1.3309494213981194e+04, + "cpu_time": 2.6028741551035906e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2310009483898447e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52610, + "real_time": 1.3301689696182539e+04, + "cpu_time": 2.5502100038015586e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2317232151868689e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52610, + "real_time": 1.3304101858946815e+04, + "cpu_time": 2.5788528948869014e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2314998918158460e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3306610874460541e+04, + "cpu_time": 2.5828995654818464e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2312677690354698e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3306710123221137e+04, + "cpu_time": 2.5821369834632205e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2312585040391598e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8281146294386743e+00, + "cpu_time": 2.1163926844984059e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5422795598314935e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8768517134486872e-04, + "cpu_time": 8.1938636437208397e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8769368036055928e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52590, + "real_time": 1.3305105829340244e+04, + "cpu_time": 2.6095209184255531e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4628139317569680e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52590, + "real_time": 1.3305670310677066e+04, + "cpu_time": 2.6090151815934612e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4627094490463581e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52590, + "real_time": 1.3309898662158424e+04, + "cpu_time": 2.6204574367750560e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4619270838750410e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52590, + "real_time": 1.3302018918613057e+04, + "cpu_time": 2.5678119680547628e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4633854605445542e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52590, + "real_time": 1.3311707294286807e+04, + "cpu_time": 2.6198353146986115e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4615925873057284e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3306880203015122e+04, + "cpu_time": 2.6053281639094890e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4624857025057302e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3305670310677064e+04, + "cpu_time": 2.6095209184255531e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4627094490463581e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8941055772511626e+00, + "cpu_time": 2.1667894237232744e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.2060235514463868e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9263850863922423e-04, + "cpu_time": 8.3167619869884100e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9263209707629229e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52589, + "real_time": 1.3311737917013636e+04, + "cpu_time": 2.6368979767631947e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.9231738491665249e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52589, + "real_time": 1.3308368375146654e+04, + "cpu_time": 2.5941617961931264e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.9244203461025562e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52589, + "real_time": 1.3305373496620627e+04, + "cpu_time": 2.5969508794614816e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.9255287735173464e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52589, + "real_time": 1.3310081896698615e+04, + "cpu_time": 2.6246957253417964e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.9237863830315952e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52589, + "real_time": 1.3306398366144516e+04, + "cpu_time": 2.6042027382152137e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.9251494053224287e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3308392010324813e+04, + "cpu_time": 2.6113818231949634e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.9244117514280910e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3308368375146654e+04, + "cpu_time": 2.6042027382152144e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.9244203461025562e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6046788777727161e+00, + "cpu_time": 1.8605475077332491e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6377293790601939e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9571702394639227e-04, + "cpu_time": 7.1247624196790716e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9571331288991483e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52594, + "real_time": 1.3313056014269650e+04, + "cpu_time": 2.6253008270905520e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.8453728324668636e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52594, + "real_time": 1.3309369376873476e+04, + "cpu_time": 2.6291280222078498e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.8480999578952503e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52594, + "real_time": 1.3307968962511117e+04, + "cpu_time": 2.6073490607293599e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.8491362858775158e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52594, + "real_time": 1.3309378249764339e+04, + "cpu_time": 2.6259484465908590e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.8480933925159740e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52594, + "real_time": 1.3306352453303974e+04, + "cpu_time": 2.6090782009354662e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.8503327985615425e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3309225011344510e+04, + "cpu_time": 2.6193609115108175e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.8482070534634304e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3309369376873476e+04, + "cpu_time": 2.6253008270905520e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.8480999578952503e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4774056622765270e+00, + "cpu_time": 1.0296763603946003e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8329935862844693e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8614199250255648e-04, + "cpu_time": 3.9310213261168920e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8612459875524650e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51681, + "real_time": 1.3551560673031219e+04, + "cpu_time": 2.6843285424043625e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9344192622897621e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51681, + "real_time": 1.3549864229279214e+04, + "cpu_time": 2.6722629844623716e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9346614516885445e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51681, + "real_time": 1.3543381228191653e+04, + "cpu_time": 2.6214132834116990e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9355875433405499e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51681, + "real_time": 1.3545087871039526e+04, + "cpu_time": 2.6354748611675444e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9353436647722656e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51681, + "real_time": 1.3549848979364944e+04, + "cpu_time": 2.6721088852769793e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9346636290870762e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3547948596181313e+04, + "cpu_time": 2.6571177113445912e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9349351102356396e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3549848979364944e+04, + "cpu_time": 2.6721088852769790e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9346636290870762e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5132892026029303e+00, + "cpu_time": 2.7100567606817447e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0181567527848314e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5932259615992360e-04, + "cpu_time": 1.0199234866830059e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5934496336539739e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51504, + "real_time": 1.3596964597759266e+04, + "cpu_time": 2.6881258814849243e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8559194313589729e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51504, + "real_time": 1.3594353967578696e+04, + "cpu_time": 2.6535145580925739e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8566599137434517e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51504, + "real_time": 1.3594027187301357e+04, + "cpu_time": 2.6773934801180501e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8567526221350746e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51504, + "real_time": 1.3595096543920763e+04, + "cpu_time": 2.6918847157502310e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8564492595269035e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51504, + "real_time": 1.3590203726458405e+04, + "cpu_time": 2.6675664783317716e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8578376789104179e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3594129204603698e+04, + "cpu_time": 2.6756970227555103e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8567237811349648e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3594353967578698e+04, + "cpu_time": 2.6773934801180498e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8566599137434517e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4721197546445413e+00, + "cpu_time": 1.5633005089627898e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.0143063818570115e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8185201254430845e-04, + "cpu_time": 5.8425916524467245e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8187214796577491e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51227, + "real_time": 1.3664408083870647e+04, + "cpu_time": 2.6949320026548547e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6737755017557648e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51227, + "real_time": 1.3664886441811936e+04, + "cpu_time": 2.7261418724500887e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6735068708039764e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51227, + "real_time": 1.3665361331699922e+04, + "cpu_time": 2.6794068459991860e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6732402060060333e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51227, + "real_time": 1.3663284735924808e+04, + "cpu_time": 2.6703129833876676e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6744064129980713e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51227, + "real_time": 1.3666151912743533e+04, + "cpu_time": 2.6894128428368040e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6727963123416962e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3664818501210169e+04, + "cpu_time": 2.6920413094657200e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6735450607811081e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3664886441811934e+04, + "cpu_time": 2.6894128428368036e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.6735068708039764e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0721643327663730e+00, + "cpu_time": 2.1271243270095147e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0209007851250963e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.8461659236192646e-05, + "cpu_time": 7.9015292950005925e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8463092839546242e-05, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50517, + "real_time": 1.3843383446172982e+04, + "cpu_time": 2.7520953916503473e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.5149128882793173e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50517, + "real_time": 1.3815204847501149e+04, + "cpu_time": 2.7252342637132224e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.5180028259800479e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50517, + "real_time": 1.3851623912503235e+04, + "cpu_time": 2.7214600253379929e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.5140116518085620e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50517, + "real_time": 1.3840612178927026e+04, + "cpu_time": 2.7079353049468486e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.5152162150695987e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50517, + "real_time": 1.3869015765419288e+04, + "cpu_time": 2.7423016826018818e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.5121130695005731e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3843968030104737e+04, + "cpu_time": 2.7298053336500590e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.5148513301276196e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3843383446172980e+04, + "cpu_time": 2.7252342637132224e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.5149128882793173e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9525148710391694e+01, + "cpu_time": 1.7477431371988735e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1372872689847943e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4103722767874651e-03, + "cpu_time": 6.4024460486416558e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4108891258686998e-03, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 42863, + "real_time": 1.6360299394772917e+04, + "cpu_time": 3.0158223432797538e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.5637085842940451e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 42863, + "real_time": 1.6343953496099784e+04, + "cpu_time": 2.9876944310944342e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.5662725979983374e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 42863, + "real_time": 1.6352270280316396e+04, + "cpu_time": 2.9848514079742152e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.5649673886865607e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 42863, + "real_time": 1.6360866513469306e+04, + "cpu_time": 2.9902120196906417e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.5636197181530588e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 42863, + "real_time": 1.6357955367613411e+04, + "cpu_time": 3.0124403144903539e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.5640759531011850e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6355069010454361e+04, + "cpu_time": 2.9982041033058798e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.5645288484466373e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6357955367613413e+04, + "cpu_time": 2.9902120196906413e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.5640759531011850e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0831943474900454e+00, + "cpu_time": 1.4711332126890659e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1110181602719840e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.3308862487601734e-04, + "cpu_time": 4.9067146931957200e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3322505845272115e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 39482, + "real_time": 1.7714926408782871e+04, + "cpu_time": 3.1706114254597062e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7353332474703467e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 39482, + "real_time": 1.7720761540200441e+04, + "cpu_time": 3.1495596727622567e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7337739865016632e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 39482, + "real_time": 1.7707510511396216e+04, + "cpu_time": 3.1494350260878302e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7373164028908813e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 39482, + "real_time": 1.7730159663621245e+04, + "cpu_time": 3.1266047717947400e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7312647822409363e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 39482, + "real_time": 1.7740599708199035e+04, + "cpu_time": 3.1926349425054250e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7284805124840857e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7722791566439966e+04, + "cpu_time": 3.1577691677219922e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7332337863175830e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7720761540200441e+04, + "cpu_time": 3.1495596727622567e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7337739865016632e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2950976059287967e+01, + "cpu_time": 2.4943030754190249e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4582474113519549e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.3075260241801021e-04, + "cpu_time": 7.8989404954460610e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3063101623012014e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 29369, + "real_time": 2.3830339467114180e+04, + "cpu_time": 3.8199409990125845e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.0402757053261975e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 29369, + "real_time": 2.3814132114149987e+04, + "cpu_time": 3.7792519561442066e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.0450671557462476e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 29369, + "real_time": 2.3840527193956808e+04, + "cpu_time": 3.7698421294562300e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.0372671977877893e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 29369, + "real_time": 2.3855154916492960e+04, + "cpu_time": 3.7879603016786357e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.0329520217873669e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 29369, + "real_time": 2.3826329507420520e+04, + "cpu_time": 3.7891303244917908e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.0414605803108984e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3833296639826895e+04, + "cpu_time": 3.7892251421566893e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.0394045321917004e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3830339467114180e+04, + "cpu_time": 3.7879603016786350e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.0402757053261975e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5448543781939838e+01, + "cpu_time": 1.8843024536868947e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5621998773926526e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.4819164614115447e-04, + "cpu_time": 4.9727909612000991e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4809457341588855e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 19202, + "real_time": 3.6402475621909871e+04, + "cpu_time": 5.0827184095406461e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 9.2176236442019092e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 19202, + "real_time": 3.6416169226308797e+04, + "cpu_time": 5.0937231590459327e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 9.2141575330110950e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 19202, + "real_time": 3.6416209841437783e+04, + "cpu_time": 5.0891631600874280e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 9.2141472564282666e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 19202, + "real_time": 3.6425779801784782e+04, + "cpu_time": 5.0753528590771544e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 9.2117264702610156e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 19202, + "real_time": 3.6428960933663802e+04, + "cpu_time": 5.1182222268513571e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 9.2109220631084570e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6417919085021007e+04, + "cpu_time": 5.0918359629205042e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 9.2137153934021484e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6416209841437783e+04, + "cpu_time": 5.0891631600874287e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 9.2141472564282666e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0346465114925795e+01, + "cpu_time": 1.6292350968912402e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6179632752704352e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8410368782387079e-04, + "cpu_time": 3.1997006752683488e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8413763215924092e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 11336, + "real_time": 6.1765670340187135e+04, + "cpu_time": 7.7302944865914891e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.0865074989129742e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 11336, + "real_time": 6.1787648030298609e+04, + "cpu_time": 7.7190223006351662e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.0861210312956410e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 11336, + "real_time": 6.1777582144092652e+04, + "cpu_time": 7.7077260673959288e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.0862980011660612e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 11336, + "real_time": 6.1828309724881154e+04, + "cpu_time": 7.7020379146082807e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.0854067384118352e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 11336, + "real_time": 6.1795989701968807e+04, + "cpu_time": 7.7107684721242025e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.0859744187875986e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1791039988285673e+04, + "cpu_time": 7.7139698482710141e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.0860615377148221e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1787648030298602e+04, + "cpu_time": 7.7107684721242025e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.0861210312956410e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3715875075265053e+01, + "cpu_time": 1.0995746234157704e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1673962915208840e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.8380766984600191e-04, + "cpu_time": 1.4254328770318253e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8371640526829503e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 6253, + "real_time": 1.1195672692946390e+05, + "cpu_time": 1.2991731456900774e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.1988357616470977e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 6253, + "real_time": 1.1198740765995756e+05, + "cpu_time": 1.3012446969454557e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.1985073215333582e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 6253, + "real_time": 1.1204254896028397e+05, + "cpu_time": 1.2998274540220528e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.1979174808632437e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 6253, + "real_time": 1.1205160766714500e+05, + "cpu_time": 1.2995767679513856e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.1978206363508911e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 6253, + "real_time": 1.1202021114788527e+05, + "cpu_time": 1.2991934655365619e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.1981563561133655e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1201170047294714e+05, + "cpu_time": 1.2998031060291070e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.1982475113015913e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1202021114788526e+05, + "cpu_time": 1.2995767679513856e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.1981563561133655e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9461614723081468e+01, + "cpu_time": 8.5125167808084697e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2219409239961433e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.5229904158639360e-04, + "cpu_time": 6.5490817350130615e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5234297456708904e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3257, + "real_time": 2.1494033421766633e+05, + "cpu_time": 2.4614442523795221e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2488835889133777e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3257, + "real_time": 2.1505758850632957e+05, + "cpu_time": 2.4670950107460620e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2482026691752820e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3257, + "real_time": 2.1483579013561120e+05, + "cpu_time": 2.4542550629413515e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2494913246557056e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3257, + "real_time": 2.1485298890437462e+05, + "cpu_time": 2.4557379735953317e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2493913041138723e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3257, + "real_time": 2.1488624562446878e+05, + "cpu_time": 2.4590949585507889e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2491979429391345e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1491458947769008e+05, + "cpu_time": 2.4595254516426110e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2490333659594744e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1488624562446881e+05, + "cpu_time": 2.4590949585507889e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2491979429391345e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9329243199071954e+01, + "cpu_time": 5.0815187841360449e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1899464934415931e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.1564997246659730e-04, + "cpu_time": 2.0660565967074327e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1551704180895226e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1588, + "real_time": 4.5323158073020860e+05, + "cpu_time": 5.2490483879092010e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.1845399456389133e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1588, + "real_time": 4.4069463674577838e+05, + "cpu_time": 5.1218194017631048e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.2182379072375742e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1588, + "real_time": 4.5350031710672018e+05, + "cpu_time": 5.2535287657430884e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.1838380079316694e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1588, + "real_time": 4.4088931983689865e+05, + "cpu_time": 5.1235949685138831e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.2176999710462673e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1588, + "real_time": 4.5392077110994543e+05, + "cpu_time": 5.2565086649873480e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.1827414521860754e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4844732510591036e+05, + "cpu_time": 5.2009000377833255e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.1974114568080999e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5323158073020854e+05, + "cpu_time": 5.2490483879092010e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.1845399456389133e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9929975985554220e+03, + "cpu_time": 7.1432112481247241e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8778233391842506e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5593799331736180e-02, + "cpu_time": 1.3734567471458708e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5682356540916206e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 717, + "real_time": 9.7282398000133515e+05, + "cpu_time": 1.2237285634588664e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1037370028630732e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 717, + "real_time": 9.7408926825921051e+05, + "cpu_time": 1.2248211562064535e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1023033093454341e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 717, + "real_time": 9.7240158868174243e+05, + "cpu_time": 1.2221209330544036e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1042164435947104e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 717, + "real_time": 9.7343945173375215e+05, + "cpu_time": 1.2234382566248472e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1030391485444766e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 717, + "real_time": 9.7216814803479123e+05, + "cpu_time": 1.2222837210599494e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1044815921716184e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7298448734216637e+05, + "cpu_time": 1.2232785260809041e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1035554993038625e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7282398000133527e+05, + "cpu_time": 1.2234382566248472e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1037370028630732e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8413595628434803e+02, + "cpu_time": 1.1110146081050393e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8910582846027970e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.0590797333914064e-04, + "cpu_time": 9.0822701814644621e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.0567386870994656e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2655056479076543e+06, + "cpu_time": 2.5682348284789855e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8079848989928272e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2724214300790555e+06, + "cpu_time": 2.5719383689320390e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8024825614575832e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2739862358577331e+06, + "cpu_time": 2.5746093042071215e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8012422130844670e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2679636696456028e+06, + "cpu_time": 2.5754732491909903e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8060254027967078e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2671512519476581e+06, + "cpu_time": 2.5689966796116554e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8066725792913991e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2694056470875405e+06, + "cpu_time": 2.5718504860841585e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8048815311245972e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time_median", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2679636696456028e+06, + "cpu_time": 2.5719383689320390e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8060254027967078e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6210609816635879e+03, + "cpu_time": 3.2388425654462471e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8786663284101314e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5955988239963639e-03, + "cpu_time": 1.2593432561383597e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5949336722485457e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2672242530196616e+06, + "cpu_time": 2.5747143236245373e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6132288145247530e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2682318762306450e+06, + "cpu_time": 2.5709937184465667e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6116236994312471e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2686626822431497e+06, + "cpu_time": 2.5824757702264967e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6109378728353418e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2691023523970521e+06, + "cpu_time": 2.5715750064725624e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6102382033785610e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2812402049157899e+06, + "cpu_time": 2.5840370129449815e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.5910291175594986e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2708922737612594e+06, + "cpu_time": 2.5767591663430287e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6074115415458810e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time_median", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2686626822431497e+06, + "cpu_time": 2.5747143236245373e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6109378728353418e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.8263768155085991e+03, + "cpu_time": 6.1225790292166594e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.2248547968490657e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5656773255291502e-03, + "cpu_time": 2.3760773258084131e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5571950110510394e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2758004594255579e+06, + "cpu_time": 2.5800106363635971e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1992251922365548e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2812633224847643e+06, + "cpu_time": 2.5926756201298591e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1819854545131857e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2750321372000910e+06, + "cpu_time": 2.5801103668831284e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2016565094170412e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2748477330624862e+06, + "cpu_time": 2.5852198538960805e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2022402914604032e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2752944768599970e+06, + "cpu_time": 2.5859965681817885e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2008261640974991e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2764476258065789e+06, + "cpu_time": 2.5848026090908903e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1971867223449368e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time_median", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2752944768599970e+06, + "cpu_time": 2.5852198538960801e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2008261640974991e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7158229003427468e+03, + "cpu_time": 5.2097042603076579e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.5731703108994334e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1930091733959795e-03, + "cpu_time": 2.0155133865869864e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1911835334607219e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 306, + "real_time": 2.3023587981774625e+06, + "cpu_time": 2.6108053921568762e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4232360319312096e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 306, + "real_time": 2.2894127432902167e+06, + "cpu_time": 2.5977911241829190e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4312840747495646e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 306, + "real_time": 2.2892052120547378e+06, + "cpu_time": 2.6040253921568999e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4314138298937473e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 306, + "real_time": 2.2956695692199799e+06, + "cpu_time": 2.6043896045751143e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4273831233967124e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 306, + "real_time": 2.2899173550137314e+06, + "cpu_time": 2.6050263790849531e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4309686735311681e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2933127355512260e+06, + "cpu_time": 2.6044075784313525e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4288571467004806e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time_median", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2899173550137309e+06, + "cpu_time": 2.6043896045751148e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4309686735311681e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.7226788282562266e+03, + "cpu_time": 4.6162428159335204e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5586103401095177e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4953765526796808e-03, + "cpu_time": 1.7724732696077878e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4905291255511910e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 299, + "real_time": 2.3224923127396251e+06, + "cpu_time": 2.6416980434782128e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8217962074842505e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 299, + "real_time": 2.3290524064535182e+06, + "cpu_time": 2.6490035050166822e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8138482336596545e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 299, + "real_time": 2.3288207119299616e+06, + "cpu_time": 2.6481698929765620e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8141281836027820e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 299, + "real_time": 2.3275712118748836e+06, + "cpu_time": 2.6536149130435167e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8156388799468800e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 299, + "real_time": 2.3217196186948381e+06, + "cpu_time": 2.6471683846153934e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8227353325653192e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3259312523385654e+06, + "cpu_time": 2.6479309478260735e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8176293674517769e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time_median", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3275712118748836e+06, + "cpu_time": 2.6481698929765616e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8156388799468800e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5476827131645537e+03, + "cpu_time": 4.2704078718644978e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2997307244901975e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5252741067035408e-03, + "cpu_time": 1.6127338499406351e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5260100473678735e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 294, + "real_time": 2.3838274630711596e+06, + "cpu_time": 2.7287894659863906e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4983845110642292e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 294, + "real_time": 2.3836039329821966e+06, + "cpu_time": 2.7295267312925630e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4989001396726169e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 294, + "real_time": 2.3893008490099390e+06, + "cpu_time": 2.7346804183673882e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4857888680830069e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 294, + "real_time": 2.3826755849378449e+06, + "cpu_time": 2.7285238095238302e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5010426441843607e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 294, + "real_time": 2.3894603245378453e+06, + "cpu_time": 2.7304244353741426e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4854227397707947e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3857736309077973e+06, + "cpu_time": 2.7303889721088633e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4939077805550016e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time_median", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3838274630711600e+06, + "cpu_time": 2.7295267312925630e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4983845110642292e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3213711260512400e+03, + "cpu_time": 2.5099043391209511e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6449934834831613e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3921568597383830e-03, + "cpu_time": 9.1924790378214254e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3915401912171985e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 280, + "real_time": 2.5049788395075924e+06, + "cpu_time": 2.8802415607142760e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0464918739654106e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 280, + "real_time": 2.5081954846557761e+06, + "cpu_time": 2.8800819178571706e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0451497963524026e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 280, + "real_time": 2.5002885243988461e+06, + "cpu_time": 2.8700542392857843e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0484549980607869e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 280, + "real_time": 2.5008531287312508e+06, + "cpu_time": 2.8761451285714451e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0482182939427260e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 280, + "real_time": 2.4995188469932014e+06, + "cpu_time": 2.8742114249999826e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0487778490462129e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5027669648573333e+06, + "cpu_time": 2.8761468542857314e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0474185622735077e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time_median", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5008531287312508e+06, + "cpu_time": 2.8761451285714451e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0482182939427260e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6985443808903647e+03, + "cpu_time": 4.2754296007076864e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5465672019831534e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4777821638305006e-03, + "cpu_time": 1.4865129693697283e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4765512639247128e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 271, + "real_time": 2.5877556875279360e+06, + "cpu_time": 3.0251080147601436e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0260336110046327e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 271, + "real_time": 2.5920514395371135e+06, + "cpu_time": 3.0290256974169980e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0226759083671078e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 271, + "real_time": 2.5885113561354112e+06, + "cpu_time": 3.0273132177121728e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0254421475003693e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 271, + "real_time": 2.5911877022856828e+06, + "cpu_time": 3.0249144022140317e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0233501399282128e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 271, + "real_time": 2.5867496735701594e+06, + "cpu_time": 3.0206817195571107e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0268215566309223e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5892511718112607e+06, + "cpu_time": 3.0254086103320918e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0248646726862490e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time_median", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5885113561354112e+06, + "cpu_time": 3.0251080147601436e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0254421475003693e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2711750009013181e+03, + "cpu_time": 3.1380305930295453e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7757947127976251e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.7715515034895650e-04, + "cpu_time": 1.0372253791811254e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7699426867959536e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 255, + "real_time": 2.7534401199469962e+06, + "cpu_time": 3.3270618627450494e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8082397085874707e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 255, + "real_time": 2.7493240193043854e+06, + "cpu_time": 3.3218148666666071e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8139411456686115e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 255, + "real_time": 2.7509195629653395e+06, + "cpu_time": 3.3211108509804299e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8117290455039442e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 255, + "real_time": 2.7490083265173086e+06, + "cpu_time": 3.3206275098039252e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8143791340509713e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 255, + "real_time": 2.7506893199375449e+06, + "cpu_time": 3.3193359921568246e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8120481015420830e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7506762697343146e+06, + "cpu_time": 3.3219902164705675e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8120674270706159e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time_median", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7506893199375449e+06, + "cpu_time": 3.3211108509804299e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8120481015420830e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7541982283939942e+03, + "cpu_time": 2.9760503744443181e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4300779271455473e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.3773343584464426e-04, + "cpu_time": 8.9586367825195113e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.3746981752968131e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 226, + "real_time": 3.0820040512351994e+06, + "cpu_time": 3.9697559867257560e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8045076032898390e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 226, + "real_time": 3.0826270679193260e+06, + "cpu_time": 3.9766436371681304e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8031323731141770e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 226, + "real_time": 3.0827565979937799e+06, + "cpu_time": 3.9774748053097669e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8028465217292881e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 226, + "real_time": 3.0776273255211720e+06, + "cpu_time": 3.9644059115043832e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8141843640696931e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 226, + "real_time": 3.0808082196564800e+06, + "cpu_time": 3.9734567831858508e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8071488079639018e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0811646524651917e+06, + "cpu_time": 3.9723474247787772e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8063639340333796e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time_median", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0820040512351994e+06, + "cpu_time": 3.9734567831858499e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8045076032898390e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1223375133371210e+03, + "cpu_time": 5.3796220209078874e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6913828502905194e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.8881015872977519e-04, + "cpu_time": 1.3542677529540313e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8926417919449321e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 193, + "real_time": 3.6185814376553725e+06, + "cpu_time": 5.0328191813470777e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1591017287475121e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 193, + "real_time": 3.6176951997792781e+06, + "cpu_time": 5.0271622487046085e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1593856774489741e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 193, + "real_time": 3.6136918252459462e+06, + "cpu_time": 5.0231052124352567e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1606700855611942e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 193, + "real_time": 3.6202867544338186e+06, + "cpu_time": 5.0531507772020241e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1585557400565505e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 193, + "real_time": 3.6116830867092689e+06, + "cpu_time": 5.0209306321244417e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1613156246833322e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6163876607647366e+06, + "cpu_time": 5.0314336103626816e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1598057712995129e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time_median", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6176951997792781e+06, + "cpu_time": 5.0271622487046076e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1593856774489741e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5751151362938463e+03, + "cpu_time": 1.2956411895234705e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1468902789717943e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.8858736165962822e-04, + "cpu_time": 2.5750934820147151e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.8886409030949436e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 143, + "real_time": 4.9302749908887427e+06, + "cpu_time": 7.6370821258742372e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7014482996389318e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 143, + "real_time": 4.9525420735036572e+06, + "cpu_time": 7.6561792657341706e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6937984322999420e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 143, + "real_time": 4.9444659404665027e+06, + "cpu_time": 7.6640829580420181e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6965650286608200e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 143, + "real_time": 4.9617740332142459e+06, + "cpu_time": 7.6777227972026290e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6906469226221182e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 143, + "real_time": 4.9411282724850662e+06, + "cpu_time": 7.6620138951047864e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6977110363056970e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9460370621116431e+06, + "cpu_time": 7.6594162083915677e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6960339439055016e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time_median", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9444659404665027e+06, + "cpu_time": 7.6620138951047882e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6965650286608200e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1884600968984976e+04, + "cpu_time": 1.4772836238523447e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0751517884073961e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4028531973658536e-03, + "cpu_time": 1.9287156927623929e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4027536730918471e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 88, + "real_time": 7.9950475536117498e+06, + "cpu_time": 1.3781671761363609e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0984510582955720e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 88, + "real_time": 7.9906820637089284e+06, + "cpu_time": 1.3790179761363804e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0995974894554551e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 88, + "real_time": 7.9878256973725827e+06, + "cpu_time": 1.3783974954545319e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1003482844547410e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 88, + "real_time": 8.0014766215092754e+06, + "cpu_time": 1.3793579079545410e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0967649839655976e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 88, + "real_time": 7.9991621049967678e+06, + "cpu_time": 1.3789241079545630e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0973716721555026e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.9948388082398605e+06, + "cpu_time": 1.3787729327272754e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0985066976653738e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time_median", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.9950475536117498e+06, + "cpu_time": 1.3789241079545628e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0984510582955720e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.6845686630151667e+03, + "cpu_time": 4.8295714914499968e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4921640229411779e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.1102980302196713e-04, + "cpu_time": 3.5028041070525552e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1105992875850098e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52, + "real_time": 1.3578145567757579e+07, + "cpu_time": 2.5060329076923110e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4712087399974375e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52, + "real_time": 1.3572594706112375e+07, + "cpu_time": 2.4997964076922912e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4722194043625917e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52, + "real_time": 1.3595579478603143e+07, + "cpu_time": 2.5079057692307808e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4680398546313009e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52, + "real_time": 1.3560729871432368e+07, + "cpu_time": 2.4959871865384832e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4743824497741265e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52, + "real_time": 1.3564868239112772e+07, + "cpu_time": 2.5004916711538367e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4736275655999055e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3574383572603649e+07, + "cpu_time": 2.5020427884615410e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4718956028730726e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time_median", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3572594706112375e+07, + "cpu_time": 2.5004916711538367e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4722194043625917e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3634370256295786e+04, + "cpu_time": 4.8584488545905326e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4813063552894876e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0044191092266764e-03, + "cpu_time": 1.9417928730059418e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0038070994606233e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 28, + "real_time": 2.4551194693361010e+07, + "cpu_time": 4.7682736892857596e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7334255965208564e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 28, + "real_time": 2.4667675540383372e+07, + "cpu_time": 4.7669700642856583e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7205183516434813e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 28, + "real_time": 2.4626092652657203e+07, + "cpu_time": 4.7664402500000082e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7251121380296936e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 28, + "real_time": 2.4654709534453493e+07, + "cpu_time": 4.7663174535713896e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7219490826375117e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 28, + "real_time": 2.4592131775404725e+07, + "cpu_time": 4.7679811071428649e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7288754229561114e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4618360839251962e+07, + "cpu_time": 4.7671965128571369e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7259761183575311e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time_median", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4626092652657203e+07, + "cpu_time": 4.7669700642856583e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7251121380296936e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7466312456408908e+04, + "cpu_time": 8.9048510839228311e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2595481806254042e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9280858204307311e-03, + "cpu_time": 1.8679429429658363e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9294182899131243e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 16, + "real_time": 4.4878440676257014e+07, + "cpu_time": 8.6938621187499627e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9906949969188218e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 16, + "real_time": 4.4930730946362019e+07, + "cpu_time": 8.6998784062499717e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9872144337074809e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 16, + "real_time": 4.5045532751828432e+07, + "cpu_time": 8.6797669312499434e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9796013011867867e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 16, + "real_time": 4.4967670692130923e+07, + "cpu_time": 8.7004491562499449e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9847605164811730e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 16, + "real_time": 4.4860868481919169e+07, + "cpu_time": 8.6980766499998197e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9918664649592199e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4936648709699519e+07, + "cpu_time": 8.6944066524999291e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9868275426506968e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time_median", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4930730946362019e+07, + "cpu_time": 8.6980766499998197e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9872144337074809e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4134904099876760e+04, + "cpu_time": 8.5810660405493269e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9241442938403012e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6497648629474862e-03, + "cpu_time": 9.8696396240932518e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6486202244774765e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.5508892312645912e+07, + "cpu_time": 1.6437865512499884e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1392694811028571e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.4170997142791748e+07, + "cpu_time": 1.6137153162500083e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1891680639664173e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.5521608591079712e+07, + "cpu_time": 1.6437616275000265e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1388027005375929e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.5558866150677204e+07, + "cpu_time": 1.6456377574999693e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1374358740011811e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.5441146045923233e+07, + "cpu_time": 1.6450875774999928e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1417586072139096e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.5240302048623562e+07, + "cpu_time": 1.6383977659999973e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1492869453643918e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time_median", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.5508892312645912e+07, + "cpu_time": 1.6437865512499884e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1392694811028571e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.9927215869783680e+05, + "cpu_time": 1.3822123544765990e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2348917023691326e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.0303852085835464e-03, + "cpu_time": 8.4363662058155025e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.0965007036236961e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6664851829409599e+08, + "cpu_time": 3.1134529749999726e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2215762701984982e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6343548521399498e+08, + "cpu_time": 3.0650404024999034e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2849103197940502e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6666215285658836e+08, + "cpu_time": 3.1133070274999851e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2213127143627729e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.6385103389620781e+08, + "cpu_time": 3.0709383099998885e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2765793369365211e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6660070046782494e+08, + "cpu_time": 3.1135498049999911e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2225009288222299e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6543957814574242e+08, + "cpu_time": 3.0952577039999479e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2453759140228143e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time_median", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6660070046782494e+08, + "cpu_time": 3.1133070274999851e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2225009288222299e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6465335985439108e+06, + "cpu_time": 2.4979817648659674e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2424328879694026e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.9524770130483085e-03, + "cpu_time": 8.0703514981575481e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9909316327865278e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.1833836436271667e+08, + "cpu_time": 5.5839567699999297e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3729576582751179e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.1884165108203888e+08, + "cpu_time": 5.5723308749998069e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3676334956743875e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.1864756345748901e+08, + "cpu_time": 5.5802271249999082e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3696847148283582e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.1875763833522797e+08, + "cpu_time": 5.5823924349999738e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3685210795506568e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.1939861178398132e+08, + "cpu_time": 5.5973558450000381e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3617610859442406e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1879676580429077e+08, + "cpu_time": 5.5832526099999309e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3681116068545523e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time_median", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1875763833522797e+08, + "cpu_time": 5.5823924349999738e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3685210795506568e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8670319761437364e+05, + "cpu_time": 9.0633166487426159e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0830214634948955e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2130085342577492e-03, + "cpu_time": 1.6233040633894459e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2122583631686691e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2671090729244342e+06, + "cpu_time": 2.5772651396102724e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8067061920035486e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2667771335328361e+06, + "cpu_time": 2.5822860681817667e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8069707601189131e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2744213453155351e+06, + "cpu_time": 2.5852891038960540e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8008976254273388e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2662420841780575e+06, + "cpu_time": 2.5826376720779641e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8073973776219836e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2668831138085430e+06, + "cpu_time": 2.5831289870129973e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8068862814538311e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2682865499518812e+06, + "cpu_time": 2.5821213941558106e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8057716473251232e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time_median", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2668831138085425e+06, + "cpu_time": 2.5826376720779641e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8068862814538311e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4441972046697256e+03, + "cpu_time": 2.9549321905544271e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7364549578028841e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5184136258017270e-03, + "cpu_time": 1.1443815915248639e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5153936888179496e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2741728062925311e+06, + "cpu_time": 2.5872905889968197e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6021888826271752e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2750763413963486e+06, + "cpu_time": 2.5894323042071145e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6007582914655455e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2682309953923845e+06, + "cpu_time": 2.5868045339806746e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6116251019587419e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2675873731296137e+06, + "cpu_time": 2.5861600517799933e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6126502101190486e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2683526723853592e+06, + "cpu_time": 2.5860219288025321e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6114313703192542e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2706840377192474e+06, + "cpu_time": 2.5871418815534269e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6077307712979536e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time_median", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2683526723853592e+06, + "cpu_time": 2.5868045339806741e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6114313703192542e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6230460356078656e+03, + "cpu_time": 1.3784263753519763e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7530237215674651e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5955747146780387e-03, + "cpu_time": 5.3279891032660036e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5946377615915334e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2763356546774604e+06, + "cpu_time": 2.5985404772728696e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1975325635012686e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2815434933434445e+06, + "cpu_time": 2.5986161396104251e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1811035151428906e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2758083824709347e+06, + "cpu_time": 2.5971933409090815e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1992001287082201e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2755501936976011e+06, + "cpu_time": 2.5978951493506762e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2000169652936589e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2836131067922362e+06, + "cpu_time": 2.5999726461039130e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1745953599882806e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2785701661963356e+06, + "cpu_time": 2.5984435506493929e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1904897065268643e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time_median", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2763356546774600e+06, + "cpu_time": 2.5985404772728691e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1975325635012686e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7420799411557246e+03, + "cpu_time": 1.0296944931755434e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1800135477406870e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6422930470482095e-03, + "cpu_time": 3.9627356650415055e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6410753591227303e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 306, + "real_time": 2.2894217483577579e+06, + "cpu_time": 2.6157123986928468e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4312784450267870e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 306, + "real_time": 2.2982097583902036e+06, + "cpu_time": 2.6196216928103981e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4258054505413188e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 306, + "real_time": 2.2948424200971727e+06, + "cpu_time": 2.6147066568628689e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4278976069569288e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 306, + "real_time": 2.2951713040337162e+06, + "cpu_time": 2.6159913267973727e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4276929980089467e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 306, + "real_time": 2.2892230433819732e+06, + "cpu_time": 2.6157642941176118e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4314026802556708e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2933736548521654e+06, + "cpu_time": 2.6163592738562198e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4288154361579306e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time_median", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2948424200971727e+06, + "cpu_time": 2.6157642941176118e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4278976069569288e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9249793588278822e+03, + "cpu_time": 1.8896000741518933e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4455334774754414e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7114434669307707e-03, + "cpu_time": 7.2222499907928733e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7115810870936940e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 301, + "real_time": 2.3333618852102580e+06, + "cpu_time": 2.6640279767441349e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8086513461710457e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 301, + "real_time": 2.3237798854931449e+06, + "cpu_time": 2.6616466777408584e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8202326911049999e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 301, + "real_time": 2.3243023961336906e+06, + "cpu_time": 2.6552495847176355e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8195986937420197e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 301, + "real_time": 2.3317873660487393e+06, + "cpu_time": 2.6638293588040257e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8105478635923855e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 301, + "real_time": 2.3322657471603137e+06, + "cpu_time": 2.6639117209302289e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8099713799679294e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3290994560092292e+06, + "cpu_time": 2.6617330637873770e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8138003949156761e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time_median", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3317873660487393e+06, + "cpu_time": 2.6638293588040262e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8105478635923855e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6563981964097811e+03, + "cpu_time": 3.7566803371265755e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6291145396056738e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9992268618654169e-03, + "cpu_time": 1.4113662967319494e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0005379734031798e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 294, + "real_time": 2.3916211687218472e+06, + "cpu_time": 2.7459815748298741e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4804666271643989e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 294, + "real_time": 2.3832140137207042e+06, + "cpu_time": 2.7419768911564848e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4997998184547730e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 294, + "real_time": 2.3888418083490967e+06, + "cpu_time": 2.7433556292517390e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4868430191525519e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 294, + "real_time": 2.3704868920312142e+06, + "cpu_time": 2.7251113231292660e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5293281916310236e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 294, + "real_time": 2.3771813091802960e+06, + "cpu_time": 2.7292443061224613e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5137569647641420e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3822690384006323e+06, + "cpu_time": 2.7371339448979655e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5020389242333785e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time_median", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3832140137207042e+06, + "cpu_time": 2.7419768911564853e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4997998184547730e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6061110030168456e+03, + "cpu_time": 9.3171062335405913e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9895412258074724e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6125688846607657e-03, + "cpu_time": 3.4039643002885321e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6160071806192887e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 280, + "real_time": 2.4899323221429116e+06, + "cpu_time": 2.8720195607143799e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0528157639818534e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 280, + "real_time": 2.5041672584068561e+06, + "cpu_time": 2.8824062892856058e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0468310338294864e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 280, + "real_time": 2.4900311121850139e+06, + "cpu_time": 2.8712969071429078e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0527739943376346e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 280, + "real_time": 2.4899225657074046e+06, + "cpu_time": 2.8706513678571358e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0528198893025535e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 280, + "real_time": 2.5060735119041055e+06, + "cpu_time": 2.8821455964286914e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0460347581776400e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4960253540692581e+06, + "cpu_time": 2.8757039442857439e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0502550879258336e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time_median", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4900311121850139e+06, + "cpu_time": 2.8720195607143799e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0527739943376346e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.3300072287625117e+03, + "cpu_time": 6.0195825867577423e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5005530675523262e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.3373087397458282e-03, + "cpu_time": 2.0932553223078263e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3330503301494375e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 270, + "real_time": 2.6007938043524823e+06, + "cpu_time": 3.0453855814814521e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0158768416111773e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 270, + "real_time": 2.5987131822923268e+06, + "cpu_time": 3.0437650592592424e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0174908241990954e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 270, + "real_time": 2.5980696263205674e+06, + "cpu_time": 3.0476056296295072e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0179905676450482e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 270, + "real_time": 2.5992518704798487e+06, + "cpu_time": 3.0439464111111644e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0170727044748113e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 270, + "real_time": 2.5948500303827501e+06, + "cpu_time": 3.0430851222223300e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0204944172541082e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5983357027655952e+06, + "cpu_time": 3.0447575607407391e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0177850710368481e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time_median", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5987131822923264e+06, + "cpu_time": 3.0439464111111648e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0174908241990954e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1934462675495224e+03, + "cpu_time": 1.7990261221939127e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7042720325699181e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.4417354740377848e-04, + "cpu_time": 5.9086022000261966e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4462515707590708e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 253, + "real_time": 2.7542117469586991e+06, + "cpu_time": 3.3392219525691005e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8071727824045330e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 253, + "real_time": 2.7562702505337629e+06, + "cpu_time": 3.3408958814227306e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8043294187024623e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 253, + "real_time": 2.7555580998185594e+06, + "cpu_time": 3.3412267470354708e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8053126155062521e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 253, + "real_time": 2.7586334953710376e+06, + "cpu_time": 3.3421837628459712e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8010703551577300e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 253, + "real_time": 2.7571941556578334e+06, + "cpu_time": 3.3403462371542566e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8030546301873416e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7563735496679791e+06, + "cpu_time": 3.3407749162055063e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8041879603916645e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time_median", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7562702505337624e+06, + "cpu_time": 3.3408958814227306e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8043294187024623e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6680654128489557e+03, + "cpu_time": 1.0953417640821144e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3020600417886584e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.0516667381671973e-04, + "cpu_time": 3.2787056642721000e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0513835429720645e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 226, + "real_time": 3.1047336390423300e+06, + "cpu_time": 4.0172321017699945e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7546921694927645e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 226, + "real_time": 3.0986635103014060e+06, + "cpu_time": 4.0056525442477488e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7679242777671289e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 226, + "real_time": 3.0971116301875594e+06, + "cpu_time": 4.0060256283186190e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7713155042880952e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 226, + "real_time": 3.0960885527650868e+06, + "cpu_time": 4.0140796769912266e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7735530307330966e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 226, + "real_time": 3.0983362987686968e+06, + "cpu_time": 4.0083765619468256e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7686390300285494e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0989867262130165e+06, + "cpu_time": 4.0102733026548834e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7672248024619269e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time_median", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0983362987686968e+06, + "cpu_time": 4.0083765619468256e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7686390300285494e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3715695858948093e+03, + "cpu_time": 5.1455111880945060e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3541911343124614e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0879587051393701e-03, + "cpu_time": 1.2830824235066653e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0867366385755347e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 192, + "real_time": 3.6207315861247480e+06, + "cpu_time": 5.0885915729169305e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1584134035434380e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 192, + "real_time": 3.6210798728764835e+06, + "cpu_time": 5.0830158541668169e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1583019837306609e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 192, + "real_time": 3.6153915377023318e+06, + "cpu_time": 5.0773906666667182e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1601244170266495e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 192, + "real_time": 3.6146815509709995e+06, + "cpu_time": 5.0784186041665794e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1603522857700419e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 192, + "real_time": 3.6168631789526748e+06, + "cpu_time": 5.0771862031249171e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1596523817675993e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6177495453254478e+06, + "cpu_time": 5.0809205802083928e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1593688943676779e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time_median", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6168631789526744e+06, + "cpu_time": 5.0784186041665794e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1596523817675993e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9892341650102326e+03, + "cpu_time": 4.8964284944730416e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.5779685578936827e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.2626896294484226e-04, + "cpu_time": 9.6368924039985984e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2613640959528466e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 142, + "real_time": 4.9907564133090870e+06, + "cpu_time": 7.7879985563378939e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6808289776735446e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 142, + "real_time": 5.4827255465772366e+06, + "cpu_time": 8.4748413591547646e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.5300069151257901e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 142, + "real_time": 4.9338850953524383e+06, + "cpu_time": 7.7094881126759257e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7002033565600870e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 142, + "real_time": 4.9333104889997290e+06, + "cpu_time": 7.7020773028167905e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7004013874060583e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 142, + "real_time": 4.9877817558885459e+06, + "cpu_time": 7.7893241901407093e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6818314053329334e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0656918600254077e+06, + "cpu_time": 7.8927459042252181e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6586544084196827e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time_median", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9877817558885459e+06, + "cpu_time": 7.7879985563378949e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6818314053329334e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3478727171014337e+05, + "cpu_time": 3.2804005479886633e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.2539967577150732e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.6348510370894477e-02, + "cpu_time": 4.1562221662711438e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3734226496443394e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 87, + "real_time": 7.8241643709954862e+06, + "cpu_time": 1.3358113022988342e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1442821500777593e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 87, + "real_time": 7.8005687836771728e+06, + "cpu_time": 1.3328027517241810e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1507682920643711e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 87, + "real_time": 7.9154310073575070e+06, + "cpu_time": 1.3513200689655066e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1195581117952182e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 87, + "real_time": 7.7943636757460823e+06, + "cpu_time": 1.3308866206896568e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1524805228432035e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 87, + "real_time": 7.8576927276692167e+06, + "cpu_time": 1.3417981701149721e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1351326122644823e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8384441130890939e+06, + "cpu_time": 1.3385237827586303e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1404443378090069e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time_median", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8241643709954862e+06, + "cpu_time": 1.3358113022988342e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1442821500777593e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9696648833239087e+04, + "cpu_time": 8.2595481725336911e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3515097576550456e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.3401164971314539e-03, + "cpu_time": 6.1706398339155220e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.3141551208870609e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.3676611542263452e+07, + "cpu_time": 2.5044000823529541e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4534170540934153e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.3645244999697395e+07, + "cpu_time": 2.5002783745098807e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4590567630514603e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.3616005670936668e+07, + "cpu_time": 2.4942149058823165e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4643373990084224e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3663645340677571e+07, + "cpu_time": 2.5083934941176046e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4557452395303507e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.3613423770841431e+07, + "cpu_time": 2.4942750411764570e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4648047812828822e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3642986264883306e+07, + "cpu_time": 2.5003123796078425e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4594722473933063e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time_median", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3645244999697397e+07, + "cpu_time": 2.5002783745098807e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4590567630514603e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8126731405086699e+04, + "cpu_time": 6.2378586663681803e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0703599150198400e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0616257217442326e-03, + "cpu_time": 2.4948317327239517e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0615641914210280e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27, + "real_time": 2.5437472733082596e+07, + "cpu_time": 4.8407711148149282e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6381891276770535e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27, + "real_time": 2.5253418350109346e+07, + "cpu_time": 4.8171635148146719e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6574170304238997e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27, + "real_time": 2.5287617066944089e+07, + "cpu_time": 4.8209218555554241e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6538231665855360e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27, + "real_time": 2.5397804531234283e+07, + "cpu_time": 4.8330569851851091e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6423096499332986e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27, + "real_time": 2.5444494560360909e+07, + "cpu_time": 4.8431632666666634e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6374610759431849e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5364161448346246e+07, + "cpu_time": 4.8310153474073596e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6458400101125946e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time_median", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5397804531234279e+07, + "cpu_time": 4.8330569851851091e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6423096499332986e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8151696683491755e+04, + "cpu_time": 1.1626124340890888e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.2054897942825407e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.4754429734651956e-03, + "cpu_time": 2.4065591816283985e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4792314573437863e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15, + "real_time": 4.6092402438322701e+07, + "cpu_time": 8.8279592866668597e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9119273654611473e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15, + "real_time": 4.6236743529637657e+07, + "cpu_time": 8.8365770466665104e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9028369593971667e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15, + "real_time": 4.6033927053213120e+07, + "cpu_time": 8.8142471733332664e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9156262911232934e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15, + "real_time": 4.5941819002230965e+07, + "cpu_time": 8.8011631266666263e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9214717857271237e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15, + "real_time": 4.5961531500021614e+07, + "cpu_time": 8.8250390066669613e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9202187920987115e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6053284704685219e+07, + "cpu_time": 8.8209971280000463e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9144162387614884e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time_median", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6033927053213120e+07, + "cpu_time": 8.8250390066669598e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9156262911232934e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1874901009681328e+05, + "cpu_time": 1.3661001052196955e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5031459375811154e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5785133646445938e-03, + "cpu_time": 1.5486912481621300e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5744935942196285e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.8145243935286999e+07, + "cpu_time": 1.6829974962499961e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0453765173884535e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.7316600605845451e+07, + "cpu_time": 1.6690424174999663e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0742774471001277e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.7381476536393166e+07, + "cpu_time": 1.6705389987500042e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0719949655256786e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.7145378813147545e+07, + "cpu_time": 1.6679776587499619e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0803177363605814e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.8227984495460987e+07, + "cpu_time": 1.6880974449999541e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0425205509915061e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7643336877226830e+07, + "cpu_time": 1.6757308032499763e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0628974434732695e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time_median", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7381476536393166e+07, + "cpu_time": 1.6705389987500042e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0719949655256786e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0423536051050387e+05, + "cpu_time": 9.1861205477866379e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7592161552909862e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.7532652050531850e-03, + "cpu_time": 5.4818593356228363e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7436338883618231e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6792491823434830e+08, + "cpu_time": 3.1232115199999553e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1970890183836064e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6955873742699623e+08, + "cpu_time": 3.1624364700000739e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1662827887660499e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6895264387130737e+08, + "cpu_time": 3.1371170925000060e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1776413774793544e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.6978977248072624e+08, + "cpu_time": 3.1646866625000316e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1619743884216766e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6958019882440567e+08, + "cpu_time": 3.1493968175000477e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1658820765737572e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6916125416755679e+08, + "cpu_time": 3.1473697125000232e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1737739299248891e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time_median", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6955873742699623e+08, + "cpu_time": 3.1493968175000477e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1662827887660499e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5838440655883634e+05, + "cpu_time": 1.7462299213359724e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4284416555158282e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.4832039718010431e-03, + "cpu_time": 5.5482198815114889e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5007668695219065e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.2960829138755798e+08, + "cpu_time": 5.7457173649999046e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2576298960194526e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.2888883352279663e+08, + "cpu_time": 5.7469970249999845e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2647560955442853e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.3022075891494751e+08, + "cpu_time": 5.7793509350000250e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2515878999495478e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.2831498980522156e+08, + "cpu_time": 5.7391187999999714e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2704623832040553e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.2915075123310089e+08, + "cpu_time": 5.7278634450000250e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2621582055560555e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2923672497272491e+08, + "cpu_time": 5.7478095139999831e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2613188960546799e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time_median", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2915075123310089e+08, + "cpu_time": 5.7457173649999046e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2621582055560555e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2187257599079411e+05, + "cpu_time": 1.9191616938668352e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1488311014542794e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1925639554657714e-03, + "cpu_time": 3.3389444956244962e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1920061574176157e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2564726892231833e+06, + "cpu_time": 2.5671146990291835e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8152225017224094e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2630964000453167e+06, + "cpu_time": 2.5819286537216394e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8099096441132517e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2571355023068711e+06, + "cpu_time": 2.5690336731393156e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8146894574179288e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2626977013032222e+06, + "cpu_time": 2.5745870582524976e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8102285593169916e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2642081460243384e+06, + "cpu_time": 2.5767448899677377e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8090209626672596e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2607220877805864e+06, + "cpu_time": 2.5738817948220745e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8118142250475683e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time_median", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2626977013032218e+06, + "cpu_time": 2.5745870582524971e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8102285593169916e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6267751879019743e+03, + "cpu_time": 5.9737868599348421e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9080659442335109e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6042552100963812e-03, + "cpu_time": 2.3209250991838163e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6050574634146947e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 310, + "real_time": 2.2619184414525665e+06, + "cpu_time": 2.5835331258063419e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6217044124452309e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 310, + "real_time": 2.2679716994565339e+06, + "cpu_time": 2.5840510967742051e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6120380170365530e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 310, + "real_time": 2.2688268834064086e+06, + "cpu_time": 2.5841463354839240e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6106765394548574e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 310, + "real_time": 2.2606055199679346e+06, + "cpu_time": 2.5819186774194879e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6238078371657687e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 310, + "real_time": 2.2671505583510282e+06, + "cpu_time": 2.5826930967742014e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6133462640250525e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2652946205268945e+06, + "cpu_time": 2.5832684664516319e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6163146140254922e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time_median", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2671505583510282e+06, + "cpu_time": 2.5835331258063419e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6133462640250525e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7574721238752454e+03, + "cpu_time": 9.4949848985916105e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0018036546837320e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6587123325270950e-03, + "cpu_time": 3.6755703179522364e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6596464343578885e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2777335509633669e+06, + "cpu_time": 2.5962177792208297e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1931152759594684e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2693854687106493e+06, + "cpu_time": 2.5946163668830968e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2195756189928213e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2695174375730385e+06, + "cpu_time": 2.5931888863636223e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2191558120481391e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2763636933833165e+06, + "cpu_time": 2.5948243571429858e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1974439091711082e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2774286327774744e+06, + "cpu_time": 2.5954646818181854e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1940783408956397e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2740857566815694e+06, + "cpu_time": 2.5948624142857436e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2046737914134357e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time_median", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2763636933833165e+06, + "cpu_time": 2.5948243571429858e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1974439091711082e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2612341467504721e+03, + "cpu_time": 1.1249915183317694e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3508584634163542e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8738229788522239e-03, + "cpu_time": 4.3354572949157009e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8749751932229239e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 306, + "real_time": 2.2903509214337841e+06, + "cpu_time": 2.6129054346406180e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4306977892927814e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 306, + "real_time": 2.2833432942884518e+06, + "cpu_time": 2.6128496241829651e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4350886299911967e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 306, + "real_time": 2.2830593557967469e+06, + "cpu_time": 2.6117844444444492e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4352671084438168e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 306, + "real_time": 2.2889924778913460e+06, + "cpu_time": 2.6112970457517141e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4315468624949949e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 306, + "real_time": 2.2826013360720347e+06, + "cpu_time": 2.6053897124182088e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4355551047029573e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2856694770964729e+06, + "cpu_time": 2.6108452522875913e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4336310989851493e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time_median", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2833432942884518e+06, + "cpu_time": 2.6117844444444492e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4350886299911967e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6944419190255990e+03, + "cpu_time": 3.1269382998769638e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3157649428867182e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6163500261283251e-03, + "cpu_time": 1.1976727832249642e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6153143891242461e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 299, + "real_time": 2.3225032215171736e+06, + "cpu_time": 2.6525224080268098e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8217829535318647e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 299, + "real_time": 2.3222104786431110e+06, + "cpu_time": 2.6554502809365834e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8221386735922959e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 299, + "real_time": 2.3262035654527587e+06, + "cpu_time": 2.6589453846153016e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8172942804016575e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 299, + "real_time": 2.3167137240361809e+06, + "cpu_time": 2.6498116053511444e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8288346255325463e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 299, + "real_time": 2.3165826528429687e+06, + "cpu_time": 2.6555968361204411e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8289946797094665e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3208427284984388e+06, + "cpu_time": 2.6544653030100563e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8238090425535664e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time_median", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3222104786431110e+06, + "cpu_time": 2.6554502809365829e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8221386735922959e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1401531617820410e+03, + "cpu_time": 3.4551925671941922e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0369785606365745e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7839007835143905e-03, + "cpu_time": 1.3016529405286041e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7837532512756749e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 293, + "real_time": 2.3902581310147615e+06, + "cpu_time": 2.7515517167235794e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4835918472267516e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 293, + "real_time": 2.3876729877443137e+06, + "cpu_time": 2.7484229488053573e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4895289544581458e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 293, + "real_time": 2.3890856920452248e+06, + "cpu_time": 2.7498379249146148e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4862829088308334e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 293, + "real_time": 2.3836385666904496e+06, + "cpu_time": 2.7379847235496254e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4988202419457510e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 293, + "real_time": 2.3870215212462195e+06, + "cpu_time": 2.7416290170649579e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4910271580446303e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3875353797481945e+06, + "cpu_time": 2.7458852662116275e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4898502221012227e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time_median", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3876729877443137e+06, + "cpu_time": 2.7484229488053573e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4895289544581458e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5142086826898658e+03, + "cpu_time": 5.8031507892350428e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7845187170152712e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0530560945886513e-03, + "cpu_time": 2.1133988592470891e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0536751428532171e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 281, + "real_time": 2.4982014458851884e+06, + "cpu_time": 2.8788468078292129e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0493309113713785e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 281, + "real_time": 2.4948790089436062e+06, + "cpu_time": 2.8802535053381077e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0507283081074072e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 281, + "real_time": 2.4953174563887068e+06, + "cpu_time": 2.8751268540924629e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0505436866513254e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 281, + "real_time": 2.5017306008094465e+06, + "cpu_time": 2.8821319964411766e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0478506355367844e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 281, + "real_time": 2.4991068124002195e+06, + "cpu_time": 2.8841845195728918e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0489507639260475e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4978470648854342e+06, + "cpu_time": 2.8801087366547706e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0494808611185887e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time_median", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4982014458851884e+06, + "cpu_time": 2.8802535053381082e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0493309113713785e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8285453940079719e+03, + "cpu_time": 3.4328618802142928e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1881546100992076e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1323933453618810e-03, + "cpu_time": 1.1919209287221363e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1321355673249854e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 270, + "real_time": 2.5942774636119050e+06, + "cpu_time": 3.0489251444445625e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0209403479535899e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 270, + "real_time": 2.5923349972193441e+06, + "cpu_time": 3.0408255185184991e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0224546617716271e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 270, + "real_time": 2.5885180690912185e+06, + "cpu_time": 3.0425246629630676e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0254368948023915e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 270, + "real_time": 2.5929482895191070e+06, + "cpu_time": 3.0437304518519188e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0219763044223124e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 270, + "real_time": 2.5919189737005918e+06, + "cpu_time": 3.0490488074075598e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0227792817591515e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5919995586284334e+06, + "cpu_time": 3.0450109170371215e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0227174981418145e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time_median", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5923349972193437e+06, + "cpu_time": 3.0437304518519188e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0224546617716271e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1406029396559488e+03, + "cpu_time": 3.7737364112057339e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6715000218366735e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.2585004018621714e-04, + "cpu_time": 1.2393178592862591e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2636355465961532e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 255, + "real_time": 2.7507422099291696e+06, + "cpu_time": 3.3429572274507796e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8119748052544713e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 255, + "real_time": 2.7450003789043892e+06, + "cpu_time": 3.3364585607844177e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8199484708942652e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 255, + "real_time": 2.7528561469094427e+06, + "cpu_time": 3.3435869529411960e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8090475638445836e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 255, + "real_time": 2.7443968070050082e+06, + "cpu_time": 3.3350743254902903e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8207885875815570e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 255, + "real_time": 2.7510164095563632e+06, + "cpu_time": 3.3404875725489762e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8115948576587969e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7488023904608744e+06, + "cpu_time": 3.3397129278431321e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8146708570467353e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time_median", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7507422099291692e+06, + "cpu_time": 3.3404875725489757e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8119748052544713e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8393417056235221e+03, + "cpu_time": 3.8157948569869041e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3299092101085547e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3967325257527165e-03, + "cpu_time": 1.1425517520307464e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3972133926739084e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 227, + "real_time": 3.0898835862851352e+06, + "cpu_time": 4.0045507885462544e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7871553779841149e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 227, + "real_time": 3.0871762992559811e+06, + "cpu_time": 3.9956696828192794e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7931073470129323e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 227, + "real_time": 3.0869303757349430e+06, + "cpu_time": 4.0072940352422846e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7936485269795096e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 227, + "real_time": 3.0731934775077300e+06, + "cpu_time": 3.9933044977974249e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8240155244007909e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 227, + "real_time": 3.0903953767590849e+06, + "cpu_time": 4.0053733303963104e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7860313789340937e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0855158231085753e+06, + "cpu_time": 4.0012384669603119e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7967916310622883e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time_median", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0871762992559816e+06, + "cpu_time": 4.0045507885462544e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7931073470129323e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0619710468133617e+03, + "cpu_time": 6.2987637798451005e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5598347829177293e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2887489326496513e-03, + "cpu_time": 1.5742035451913939e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2949574852185644e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 194, + "real_time": 3.6141232229750026e+06, + "cpu_time": 5.0804254845361682e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1605315428474560e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 194, + "real_time": 3.6108860813234886e+06, + "cpu_time": 5.0735906082473826e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1615719536803203e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 194, + "real_time": 3.6080671371123996e+06, + "cpu_time": 5.0640611288661873e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1624794774070575e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 194, + "real_time": 3.6080852690988132e+06, + "cpu_time": 5.0661638659792133e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1624736355101733e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 194, + "real_time": 3.6202304679708383e+06, + "cpu_time": 5.0818327061855616e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1585737529994695e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6122784356961087e+06, + "cpu_time": 5.0732147587629026e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1611260724888954e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time_median", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6108860813234881e+06, + "cpu_time": 5.0735906082473826e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1615719536803203e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0964298769279603e+03, + "cpu_time": 8.0608448082860414e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6366307826507480e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4108629685258042e-03, + "cpu_time": 1.5889027355608476e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4095203108673629e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 144, + "real_time": 4.8802905496106381e+06, + "cpu_time": 7.5373389027776718e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7188747093488655e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 144, + "real_time": 4.9177496952729095e+06, + "cpu_time": 7.5858980069444859e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7057818148132641e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 144, + "real_time": 4.9093407675779117e+06, + "cpu_time": 7.5791113888890035e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7087035504644001e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 144, + "real_time": 4.9012917863567257e+06, + "cpu_time": 7.5752886597225051e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7115096112723987e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 144, + "real_time": 4.9707423022482544e+06, + "cpu_time": 7.6681674513889356e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6875966384750731e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9158830202132892e+06, + "cpu_time": 7.5891608819445223e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7064932648748004e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time_median", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9093407675779117e+06, + "cpu_time": 7.5791113888890045e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7087035504644001e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3673262537823219e+04, + "cpu_time": 4.8040693046208638e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1629035980159318e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.8498909350292510e-03, + "cpu_time": 6.3301719114300138e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8145806488210782e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 87, + "real_time": 7.8259912137498800e+06, + "cpu_time": 1.3335225758620623e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1437816043702247e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 87, + "real_time": 7.9808391405847566e+06, + "cpu_time": 1.3629688770115327e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1021869636092844e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 87, + "real_time": 7.8105944337259075e+06, + "cpu_time": 1.3303861666667212e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1480075738610234e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 87, + "real_time": 7.8407003512156419e+06, + "cpu_time": 1.3351152724137407e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1397598745625851e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 87, + "real_time": 7.8342597738936031e+06, + "cpu_time": 1.3329061666666050e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1415189799944272e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8584769826339576e+06, + "cpu_time": 1.3389798117241325e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1350509992795091e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time_median", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8342597738936022e+06, + "cpu_time": 1.3335225758620625e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1415189799944272e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9323594814586919e+04, + "cpu_time": 1.3517870027842367e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8628938587769300e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.8215051043327546e-03, + "cpu_time": 1.0095648873477884e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7252897443928935e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.3672612917919954e+07, + "cpu_time": 2.4968078117648046e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4541345682376499e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.3693165381019022e+07, + "cpu_time": 2.5002986901959762e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4504510875558372e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.3715000750095237e+07, + "cpu_time": 2.5056506823529266e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4465497750532022e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3691515113939257e+07, + "cpu_time": 2.4982934392155316e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4507464455733185e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.3652631830351025e+07, + "cpu_time": 2.4959067862745740e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4577262770248799e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3684985198664900e+07, + "cpu_time": 2.4993914819607627e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4519216306889777e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time_median", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3691515113939255e+07, + "cpu_time": 2.4982934392155316e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4507464455733185e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3506925982289722e+04, + "cpu_time": 3.8737827092040658e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2128197745073307e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7177165806933458e-03, + "cpu_time": 1.5498903381734735e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7181706469645809e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27, + "real_time": 2.5427313314543828e+07, + "cpu_time": 4.8171502370370977e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6392432094512830e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27, + "real_time": 2.4921568179572072e+07, + "cpu_time": 4.6943311333334513e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6928026164504519e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27, + "real_time": 2.5441998270926651e+07, + "cpu_time": 4.8323659814812988e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6377198553891635e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27, + "real_time": 2.5248313943545025e+07, + "cpu_time": 4.8132400407410011e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6579542756817245e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27, + "real_time": 2.5356415176281221e+07, + "cpu_time": 4.8299337222222984e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6466227001510315e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5279121776973762e+07, + "cpu_time": 4.7974042229630291e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6548685314247313e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time_median", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5356415176281221e+07, + "cpu_time": 4.8171502370370977e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6466227001510315e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1404289948572969e+05, + "cpu_time": 5.8191994473077171e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2664609715850834e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.4671809952154686e-03, + "cpu_time": 1.2129891868302050e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.5369988937598739e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15, + "real_time": 4.5987235009670258e+07, + "cpu_time": 8.8157721133332714e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9185866028209028e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15, + "real_time": 4.5855277031660080e+07, + "cpu_time": 8.8001038000000641e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9269854352276926e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15, + "real_time": 4.5749730120102562e+07, + "cpu_time": 8.7897971399994880e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9337381367638788e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15, + "real_time": 4.5909112940231957e+07, + "cpu_time": 8.8082366466665015e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9235530683142371e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15, + "real_time": 4.7342442969481148e+07, + "cpu_time": 9.0707495200005427e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8350401792007685e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6168759614229202e+07, + "cpu_time": 8.8569318439999729e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9075806844654961e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time_median", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5909112940231957e+07, + "cpu_time": 8.8082366466665000e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9235530683142371e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6176983202130289e+05, + "cpu_time": 1.1991647020776926e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0924103079999201e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4333714779232351e-02, + "cpu_time": 1.3539278874433847e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4074967308266573e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.7171311490237713e+07, + "cpu_time": 1.6687689924999917e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0794013696818361e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.7371034547686577e+07, + "cpu_time": 1.6698298187499461e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0723621093611932e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.7616494856774807e+07, + "cpu_time": 1.6822784099998954e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0637547922774916e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.6928851902484894e+07, + "cpu_time": 1.6639381837499911e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0879903521689863e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.7251216173171997e+07, + "cpu_time": 1.6690378212500435e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0765812532311559e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7267781794071198e+07, + "cpu_time": 1.6707706452499738e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0760179753441334e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time_median", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7251216173171997e+07, + "cpu_time": 1.6690378212500435e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0765812532311559e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5333591433128351e+05, + "cpu_time": 6.8375000353771984e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9282341625206042e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9029718542530280e-03, + "cpu_time": 4.0924228916855319e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9025299052492523e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6968788579106331e+08, + "cpu_time": 3.1570625425001711e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1638729511961102e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6865859553217888e+08, + "cpu_time": 3.1316388474999714e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1831814459617553e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6979648545384407e+08, + "cpu_time": 3.1632591599998248e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1618493784780846e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.6849016025662422e+08, + "cpu_time": 3.1315797100000966e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1863635905046439e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6925453022122383e+08, + "cpu_time": 3.1552646475000757e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1719736617878637e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6917753145098689e+08, + "cpu_time": 3.1477609815000278e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1734482055856919e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time_median", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6925453022122383e+08, + "cpu_time": 3.1552646475000757e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1719736617878637e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.8976013499814342e+05, + "cpu_time": 1.5039733453094638e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1067245616213638e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.4860429156279844e-03, + "cpu_time": 4.7779146960286777e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4874511569887327e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.8685695827007294e+08, + "cpu_time": 6.2776363250003445e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.7755525680641847e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.1934462487697601e+08, + "cpu_time": 5.5911707400002801e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3623294095325613e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.1974454224109650e+08, + "cpu_time": 5.6024478450001431e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3581240088544436e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.2095086574554443e+08, + "cpu_time": 5.6141737800004423e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3455021892705584e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.2025869190692902e+08, + "cpu_time": 5.6055897249996173e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3527328098624787e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3343113660812378e+08, + "cpu_time": 5.7382036830001664e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2388481971168456e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time_median", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2025869190692902e+08, + "cpu_time": 5.6055897249996173e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3527328098624787e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9871983356412269e+07, + "cpu_time": 3.0166446087841026e+07, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5906656868803307e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.9589663581779791e-02, + "cpu_time": 5.2571236146969222e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.9987252542014373e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 310, + "real_time": 2.2557357483873926e+06, + "cpu_time": 2.5737172322582440e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8158155284492865e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 310, + "real_time": 2.2594983343996350e+06, + "cpu_time": 2.5808855419357666e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8127917766703453e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 310, + "real_time": 2.2586957536517614e+06, + "cpu_time": 2.5764592548383796e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8134359146767620e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 310, + "real_time": 2.2613507094642809e+06, + "cpu_time": 2.5788257129031606e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8113068365987123e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 310, + "real_time": 2.2568093139618156e+06, + "cpu_time": 2.5781528483871478e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8149517438890289e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2584179719729768e+06, + "cpu_time": 2.5776081180645400e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8136603600568268e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time_median", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2586957536517614e+06, + "cpu_time": 2.5781528483871478e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8134359146767620e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2151410105829741e+03, + "cpu_time": 2.6916377274049946e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7787852950411109e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.8083748804381177e-04, + "cpu_time": 1.0442385359284469e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.8077089526584622e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 310, + "real_time": 2.2610529731478421e+06, + "cpu_time": 2.5776946612902596e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6230907003452829e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 310, + "real_time": 2.2641837837234619e+06, + "cpu_time": 2.5861608935484113e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6180808549596681e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 310, + "real_time": 2.2574678746863240e+06, + "cpu_time": 2.5799723516126596e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6288445527218333e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 310, + "real_time": 2.2585244218428289e+06, + "cpu_time": 2.5796962193548861e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6271469640854220e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 310, + "real_time": 2.2619623773460905e+06, + "cpu_time": 2.5787310612900588e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6216340652011586e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2606382861493095e+06, + "cpu_time": 2.5804510374192549e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6237594274626733e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time_median", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2610529731478421e+06, + "cpu_time": 2.5796962193548861e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6230907003452829e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6933096276830220e+03, + "cpu_time": 3.3150141277378580e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3169563826459917e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1913934414827191e-03, + "cpu_time": 1.2846646108244895e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1912922115993470e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2688063282013228e+06, + "cpu_time": 2.5932572103559538e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2214185037949011e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2704814746071701e+06, + "cpu_time": 2.5907870453074411e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2160905883782627e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2695199428216633e+06, + "cpu_time": 2.5890686763752401e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2191478430588245e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2654163622535453e+06, + "cpu_time": 2.5907805307444278e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2322246245727008e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2686484727924117e+06, + "cpu_time": 2.5882891650484405e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2219209791605230e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2685745161352227e+06, + "cpu_time": 2.5904365255663004e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2221605077930428e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time_median", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2688063282013228e+06, + "cpu_time": 2.5907805307444273e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2214185037949011e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9081053820492114e+03, + "cpu_time": 1.9157005853716407e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0789485311310727e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.4110324279754677e-04, + "cpu_time": 7.3952809361072663e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4170775830468014e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 307, + "real_time": 2.2879209238975174e+06, + "cpu_time": 2.6141247166125593e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4322173313655913e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 307, + "real_time": 2.2899472907015672e+06, + "cpu_time": 2.6168503420193871e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4309499669732977e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 307, + "real_time": 2.2837015434626355e+06, + "cpu_time": 2.6101289869706132e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4348635045504197e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 307, + "real_time": 2.2893906980510256e+06, + "cpu_time": 2.6207510684036654e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4312978570191462e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 307, + "real_time": 2.2824022885569725e+06, + "cpu_time": 2.6085248892505700e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4356802989676837e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2866725489339437e+06, + "cpu_time": 2.6140760006513591e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4330017917752275e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time_median", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2879209238975174e+06, + "cpu_time": 2.6141247166125597e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4322173313655913e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4180661753574336e+03, + "cpu_time": 4.9661259437143008e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1429612715418178e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4947772810544957e-03, + "cpu_time": 1.8997634125698230e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4954351654278674e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 299, + "real_time": 2.3308925230965568e+06, + "cpu_time": 2.6657853812711225e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8116268489692688e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 299, + "real_time": 2.3229887902138624e+06, + "cpu_time": 2.6589355384614011e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8211931231044181e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 299, + "real_time": 2.3283207100694594e+06, + "cpu_time": 2.6690462709031487e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8147325115724672e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 299, + "real_time": 2.3232750730109354e+06, + "cpu_time": 2.6587875685619572e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8208454849500950e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 299, + "real_time": 2.3282891424391479e+06, + "cpu_time": 2.6631641304348223e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8147706745453261e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3267532477659923e+06, + "cpu_time": 2.6631437779264906e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8166337286283154e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time_median", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3282891424391484e+06, + "cpu_time": 2.6631641304348223e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8147706745453261e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4719717406486561e+03, + "cpu_time": 4.4301169627904919e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2036118407641421e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4921959361106443e-03, + "cpu_time": 1.6634914718122195e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4924240230593550e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 291, + "real_time": 2.3880969337747782e+06, + "cpu_time": 2.7493104467353499e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4885544278480880e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 291, + "real_time": 2.3980329086831245e+06, + "cpu_time": 2.7551501958762938e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4658132307274282e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 291, + "real_time": 2.3929481226566834e+06, + "cpu_time": 2.7538620481101535e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4774275613832407e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 291, + "real_time": 2.3912678183059613e+06, + "cpu_time": 2.7493609140892532e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4812764591485590e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 291, + "real_time": 2.3924724952577008e+06, + "cpu_time": 2.7539792096219864e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4785164828355454e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3925636557356501e+06, + "cpu_time": 2.7523325628866078e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4783176323885731e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time_median", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3924724952577013e+06, + "cpu_time": 2.7538620481101540e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4785164828355454e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5951235224781549e+03, + "cpu_time": 2.7818028680039761e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2270660908066129e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5026239798718124e-03, + "cpu_time": 1.0107073925276168e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5017504721097326e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 278, + "real_time": 2.5094039570674193e+06, + "cpu_time": 2.8925093165468797e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0446464757565418e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 278, + "real_time": 2.5126292778969668e+06, + "cpu_time": 2.8997292266183603e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0433055218532301e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 278, + "real_time": 2.5117538899531681e+06, + "cpu_time": 2.9021617014388223e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0436691311539590e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 278, + "real_time": 2.5077187498066602e+06, + "cpu_time": 2.8902813992808270e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0453484866283779e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 278, + "real_time": 2.5128105541409776e+06, + "cpu_time": 2.8962083741007256e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0432302569248632e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5108632857730384e+06, + "cpu_time": 2.8961780035971231e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0440399744633944e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time_median", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5117538899531681e+06, + "cpu_time": 2.8962083741007252e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0436691311539590e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2204450452323645e+03, + "cpu_time": 4.9154337117414680e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.2366603718551865e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.8433530324560829e-04, + "cpu_time": 1.6972139508125469e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8470370845738502e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 268, + "real_time": 2.6076766653611921e+06, + "cpu_time": 3.0645877089549270e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0105560131910765e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 268, + "real_time": 2.6103679903908009e+06, + "cpu_time": 3.0670256753730983e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0084831024973929e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 268, + "real_time": 2.6053476917084805e+06, + "cpu_time": 3.0616844626866053e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0123532903824952e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 268, + "real_time": 2.6058336831073263e+06, + "cpu_time": 3.0590759589552078e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0119779838551047e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 268, + "real_time": 2.6069095135847135e+06, + "cpu_time": 3.0605062089551808e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0111476722452909e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6072271088305023e+06, + "cpu_time": 3.0625760029850034e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0109036124342719e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time_median", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6069095135847135e+06, + "cpu_time": 3.0616844626866048e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0111476722452909e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9774951803255785e+03, + "cpu_time": 3.2084164675208399e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5243932307642934e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5846679164539828e-04, + "cpu_time": 1.0476201943702588e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5806379845275629e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 250, + "real_time": 2.7833141256123781e+06, + "cpu_time": 3.3724455359997591e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7673649206565744e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 250, + "real_time": 2.7916020993143320e+06, + "cpu_time": 3.3871314680000069e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7561800095276803e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 250, + "real_time": 2.7848885180428624e+06, + "cpu_time": 3.3729553680000212e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7652351008179975e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 250, + "real_time": 2.7864680057391524e+06, + "cpu_time": 3.3784324799999013e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7631008066136020e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 250, + "real_time": 2.7845307402312756e+06, + "cpu_time": 3.3740853760000393e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7657188870283693e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7861606977880010e+06, + "cpu_time": 3.3770100455999458e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7635199449288452e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time_median", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7848885180428624e+06, + "cpu_time": 3.3740853760000393e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7652351008179975e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2435114007936713e+03, + "cpu_time": 6.1297504358417764e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3763932935694890e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1641508701808810e-03, + "cpu_time": 1.8151413093450807e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1628457820361655e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 226, + "real_time": 3.1013178414525054e+06, + "cpu_time": 3.9722275752213309e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7621318007760108e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 226, + "real_time": 3.1019932367896610e+06, + "cpu_time": 3.9747942876106892e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7606594854165471e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 226, + "real_time": 3.1060740151813468e+06, + "cpu_time": 3.9795687699115584e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7517772910429466e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 226, + "real_time": 3.0998127026054845e+06, + "cpu_time": 3.9712687389379605e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7654152079487956e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 226, + "real_time": 3.1017292047793097e+06, + "cpu_time": 3.9729600486727566e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7612349806959176e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1021854001616617e+06, + "cpu_time": 3.9741638840708598e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7602437531760430e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time_median", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1017292047793092e+06, + "cpu_time": 3.9729600486727571e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7612349806959176e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3317761231727372e+03, + "cpu_time": 3.2859768919740209e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0780167041867832e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5165595294569539e-04, + "cpu_time": 8.2683477275428627e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5115881757977594e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 185, + "real_time": 3.7742542065176610e+06, + "cpu_time": 5.2662236324325735e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1112934557394056e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 185, + "real_time": 3.7756216234049280e+06, + "cpu_time": 5.2741234054055158e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1108909785873871e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 185, + "real_time": 3.7750317262032549e+06, + "cpu_time": 5.2759846918919524e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1110645695734136e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 185, + "real_time": 3.7680669989738916e+06, + "cpu_time": 5.2761789243240366e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1131182118423531e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 185, + "real_time": 3.7676483732522340e+06, + "cpu_time": 5.2816254270270737e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1132418910896075e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7721245856703944e+06, + "cpu_time": 5.2748272162162308e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1119218213664334e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time_median", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7742542065176615e+06, + "cpu_time": 5.2759846918919533e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1112934557394056e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9279928512944302e+03, + "cpu_time": 5.5655084792592288e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1582632557411117e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0413210810205342e-03, + "cpu_time": 1.0551072577599065e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0416768818492380e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 134, + "real_time": 5.1886738840935389e+06, + "cpu_time": 8.0066184626866067e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6167152122850151e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 134, + "real_time": 5.2337987911400953e+06, + "cpu_time": 8.0660813805971928e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6027761736275463e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 134, + "real_time": 5.2881904933323609e+06, + "cpu_time": 8.1433110223877095e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.5862908135735304e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 134, + "real_time": 5.3593276290973620e+06, + "cpu_time": 8.2071629776123706e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.5652351527187452e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 134, + "real_time": 5.2725978286019461e+06, + "cpu_time": 8.1279364552240474e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.5909819547576377e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2685177252530595e+06, + "cpu_time": 8.1102220597015843e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.5923998613924952e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time_median", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2725978286019452e+06, + "cpu_time": 8.1279364552240465e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.5909819547576377e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3699582103226800e+04, + "cpu_time": 7.6629507521577369e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9210390849937435e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2090607913854393e-02, + "cpu_time": 9.4485091724352804e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2063798368544603e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 83, + "real_time": 8.4043171548520233e+06, + "cpu_time": 1.3911874373493230e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.9962616463508987e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 83, + "real_time": 8.4143913160246536e+06, + "cpu_time": 1.3920589168674543e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.9938716147000315e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 83, + "real_time": 8.3965637934315633e+06, + "cpu_time": 1.3907235469879109e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.9981049882720387e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 83, + "real_time": 8.4124885679009445e+06, + "cpu_time": 1.3913774361445708e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.9943225912978795e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 83, + "real_time": 8.4356020417917203e+06, + "cpu_time": 1.3941384313252784e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.9888581652953987e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.4126725748001821e+06, + "cpu_time": 1.3918971537349071e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.9942838011832495e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time_median", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.4124885679009445e+06, + "cpu_time": 1.3913774361445706e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.9943225912978795e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4646281661711924e+04, + "cpu_time": 1.3417012584015101e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4685108272456061e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7409784502471026e-03, + "cpu_time": 9.6393706589692658e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7392262952683401e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 48, + "real_time": 1.4164700347464532e+07, + "cpu_time": 2.5213454604165498e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3688769389326482e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 48, + "real_time": 1.4471514732576907e+07, + "cpu_time": 2.5283794437501211e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3186537567118273e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 48, + "real_time": 1.4460541521354267e+07, + "cpu_time": 2.5265395520833064e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3204132397427354e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 48, + "real_time": 1.4504024855947742e+07, + "cpu_time": 2.5246892937500339e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3134565979621964e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 48, + "real_time": 1.4425537995217988e+07, + "cpu_time": 2.5257810479165718e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3260437157437849e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4405263890512288e+07, + "cpu_time": 2.5253469595833175e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3294888498186388e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time_median", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4460541521354269e+07, + "cpu_time": 2.5257810479165722e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3204132397427354e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3736846108612648e+05, + "cpu_time": 2.6101336029888618e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2472703182476290e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.5359906024770017e-03, + "cpu_time": 1.0335742552459147e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6470533371412742e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 26, + "real_time": 2.6563442455461390e+07, + "cpu_time": 4.8400259730770431e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5263617135663285e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 26, + "real_time": 2.6636882541844480e+07, + "cpu_time": 4.8559409653845564e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5193963255488768e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 26, + "real_time": 2.6616076508966777e+07, + "cpu_time": 4.8566216692305572e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5213657609299207e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 26, + "real_time": 2.6523620964815982e+07, + "cpu_time": 4.8404630230768710e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5301546907573819e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 26, + "real_time": 2.6596519451301832e+07, + "cpu_time": 4.8481061230771817e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5232197815536046e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6587308384478092e+07, + "cpu_time": 4.8482315507692426e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5240996544712224e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time_median", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6596519451301832e+07, + "cpu_time": 4.8481061230771825e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5232197815536046e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4709383266557648e+04, + "cpu_time": 8.0238148869174503e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2470586580361705e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6816062242938207e-03, + "cpu_time": 1.6549982819290790e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6826033990032354e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 14, + "real_time": 4.9688668123313360e+07, + "cpu_time": 9.0970973500004336e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.7011737901066132e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 14, + "real_time": 4.9784063494631223e+07, + "cpu_time": 9.1046774071425527e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.6959978470715675e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 14, + "real_time": 5.0541587173938751e+07, + "cpu_time": 9.1928049428572327e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.6555898915103321e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 14, + "real_time": 4.9743799758808956e+07, + "cpu_time": 9.1067133785713837e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.6981800475793338e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 14, + "real_time": 4.9871435388922691e+07, + "cpu_time": 9.1090056785714880e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.6912746134796848e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9925910787923001e+07, + "cpu_time": 9.1220597514286190e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.6884432379495068e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time_median", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9784063494631223e+07, + "cpu_time": 9.1067133785713851e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.6959978470715675e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5056887513064337e+05, + "cpu_time": 3.9799715659350564e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8716994941846836e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.0217822689264877e-03, + "cpu_time": 4.3630185225565395e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9620197583648491e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 7, + "real_time": 9.5718240099293843e+07, + "cpu_time": 1.7369690885714459e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8044336766068516e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 7, + "real_time": 9.6884803048201963e+07, + "cpu_time": 1.7476988785714248e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.7706662712257199e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 7, + "real_time": 9.6195612634931296e+07, + "cpu_time": 1.7440097328570962e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.7905166217791066e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 7, + "real_time": 9.5918789505958557e+07, + "cpu_time": 1.7401223585714629e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.7985700964598241e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 7, + "real_time": 9.5920642571789876e+07, + "cpu_time": 1.7392209828571075e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.7985160316153522e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6127617572035104e+07, + "cpu_time": 1.7416042082857075e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.7925405395373707e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time_median", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5920642571789891e+07, + "cpu_time": 1.7401223585714629e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.7985160316153522e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5606767617256287e+05, + "cpu_time": 4.2512213286297384e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3191745798787968e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.7443979960368789e-03, + "cpu_time": 2.4409801655304289e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7239227549310328e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.8797419592738152e+08, + "cpu_time": 3.3399792100001943e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8560883548474112e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.8609407916665077e+08, + "cpu_time": 3.3069588974998963e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8849435425574284e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.8805139139294624e+08, + "cpu_time": 3.3447744750000650e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8549159249674015e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.8614443764090538e+08, + "cpu_time": 3.3121469475000256e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8841630660792961e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.8789599463343620e+08, + "cpu_time": 3.3403039449999028e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8572770433312020e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8723201975226402e+08, + "cpu_time": 3.3288326950000167e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8674775863565483e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time_median", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8789599463343620e+08, + "cpu_time": 3.3399792100001943e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8572770433312020e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0174479874416844e+06, + "cpu_time": 1.7796431590962030e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5612702110904442e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.4341559140788000e-03, + "cpu_time": 5.3461478006067058e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4447512284628279e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.7224625051021576e+08, + "cpu_time": 6.1585569449999869e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8844933227085190e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.7160824239253998e+08, + "cpu_time": 6.1640326650001502e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8894456621491647e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.7108102440834045e+08, + "cpu_time": 6.1637449250002873e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8935508780380163e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.6749900877475739e+08, + "cpu_time": 6.1308640199996483e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.9217543404534831e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.7075382471084595e+08, + "cpu_time": 6.1606736649997628e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8961045104185247e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7063767015933990e+08, + "cpu_time": 6.1555744439999676e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8970697427535419e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time_median", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7108102440834045e+08, + "cpu_time": 6.1606736649997628e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.8935508780380163e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8432845251193587e+06, + "cpu_time": 1.3998220275730046e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4482199350849994e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.9732789555009798e-03, + "cpu_time": 2.2740721281300645e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9989129143592093e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 311, + "real_time": 2.2516541978062829e+06, + "cpu_time": 2.5739511736333407e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8191070387231780e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 311, + "real_time": 2.2594752104599569e+06, + "cpu_time": 2.5869287556271888e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8128103291587720e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 311, + "real_time": 2.2508423817358981e+06, + "cpu_time": 2.5728422765273373e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8197631398965735e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 311, + "real_time": 2.2507605728336950e+06, + "cpu_time": 2.5792025369775929e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8198292832378699e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 311, + "real_time": 2.2590610751181459e+06, + "cpu_time": 2.5810260482312427e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8131426569712309e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2543586875907956e+06, + "cpu_time": 2.5787901581993410e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8169304895975247e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time_median", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2516541978062829e+06, + "cpu_time": 2.5792025369775924e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8191070387231780e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4976759662320101e+03, + "cpu_time": 5.7052409368087438e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6224189579693380e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9951021951340932e-03, + "cpu_time": 2.2123711456974345e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9937025542302138e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 311, + "real_time": 2.2651694020274370e+06, + "cpu_time": 2.5884518360131970e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6165065591420056e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 311, + "real_time": 2.2517293564364361e+06, + "cpu_time": 2.5759160160771771e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6380926404781500e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 311, + "real_time": 2.2513584280236838e+06, + "cpu_time": 2.5815215755625935e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6386920438924539e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 311, + "real_time": 2.2573159929736252e+06, + "cpu_time": 2.5816985627011419e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6290887166437209e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 311, + "real_time": 2.2512349566606465e+06, + "cpu_time": 2.5813883344052760e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6388916118073906e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2553616272243657e+06, + "cpu_time": 2.5817952649518773e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6322543143927441e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time_median", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2517293564364361e+06, + "cpu_time": 2.5815215755625935e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6380926404781500e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.0469082372718840e+03, + "cpu_time": 4.4476217041126201e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.7159267971631016e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6811257956506551e-03, + "cpu_time": 1.7226856693439323e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6749026792160205e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2607890498737013e+06, + "cpu_time": 2.5879452750807023e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2470273159343600e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2705376824864387e+06, + "cpu_time": 2.6005925177992950e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2159119517708579e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2663341881858008e+06, + "cpu_time": 2.5942505987055358e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2292956993758194e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2627711861204463e+06, + "cpu_time": 2.5982838543688385e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2406790843446273e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2608496318043820e+06, + "cpu_time": 2.5969637766989842e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2468331239366606e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2642563476941539e+06, + "cpu_time": 2.5956072045306703e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2359494350724658e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time_median", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2627711861204463e+06, + "cpu_time": 2.5969637766989837e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2406790843446273e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1715153153597093e+03, + "cpu_time": 4.8580766012458489e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3316499592845335e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8423334970918144e-03, + "cpu_time": 1.8716532273319339e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8403251311155651e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 307, + "real_time": 2.2768980272854483e+06, + "cpu_time": 2.6145778729639226e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4391509679977411e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 307, + "real_time": 2.2762304002036876e+06, + "cpu_time": 2.6140338273615967e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4395730764806483e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 307, + "real_time": 2.2766796693035951e+06, + "cpu_time": 2.6081632410424282e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4392889980004644e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 307, + "real_time": 2.2819586521384674e+06, + "cpu_time": 2.6201082019544733e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4359594101011636e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 307, + "real_time": 2.2768463927293834e+06, + "cpu_time": 2.6091839055375601e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4391836052110288e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2777226283321166e+06, + "cpu_time": 2.6132134097719966e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4386312115582094e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time_median", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2768463927293834e+06, + "cpu_time": 2.6140338273615967e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4391836052110288e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3825491455637316e+03, + "cpu_time": 4.7909840490288143e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5027974800403495e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0460225121038443e-03, + "cpu_time": 1.8333688443175520e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0446023052792248e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 304, + "real_time": 2.3029978527981592e+06, + "cpu_time": 2.6508730526313535e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8456822015866529e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 304, + "real_time": 2.3033983820105754e+06, + "cpu_time": 2.6470629769739006e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8451873766966596e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 304, + "real_time": 2.3089810243905769e+06, + "cpu_time": 2.6574321315792594e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8383082973710150e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 304, + "real_time": 2.3029021004300662e+06, + "cpu_time": 2.6461227335528145e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8458005222089630e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 304, + "real_time": 2.3182584179602074e+06, + "cpu_time": 2.6613117631577821e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8269497262373328e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3073075555179170e+06, + "cpu_time": 2.6525605315790223e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8403856248201251e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time_median", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3033983820105754e+06, + "cpu_time": 2.6508730526313530e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8451873766966596e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6329687145745365e+03, + "cpu_time": 6.6112116131133735e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1435606510028825e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8747657410091768e-03, + "cpu_time": 2.4923885937403424e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8670616341112468e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 296, + "real_time": 2.3621171539825564e+06, + "cpu_time": 2.7292035472973529e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5489203733612925e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 296, + "real_time": 2.3686960921622813e+06, + "cpu_time": 2.7309713614865383e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5335085169305101e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 296, + "real_time": 2.3626209161168821e+06, + "cpu_time": 2.7337253716218676e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5477372229238190e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 296, + "real_time": 2.3620485538943997e+06, + "cpu_time": 2.7242265033783503e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5490815285696216e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 296, + "real_time": 2.3623361264833729e+06, + "cpu_time": 2.7337381925678058e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5484060261617705e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3635637685278985e+06, + "cpu_time": 2.7303729952703831e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5455307335894026e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time_median", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3623361264833729e+06, + "cpu_time": 2.7309713614865383e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5484060261617705e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8777330738876008e+03, + "cpu_time": 3.9391621356645142e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7410570146676546e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2175398490221158e-03, + "cpu_time": 1.4427194169031206e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2155837445525137e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 281, + "real_time": 2.4908162053555143e+06, + "cpu_time": 2.8910274341640072e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0524421650877455e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 281, + "real_time": 2.4925420562872575e+06, + "cpu_time": 2.8892442028470030e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0517134478784847e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 281, + "real_time": 2.4970106861729743e+06, + "cpu_time": 2.8889119039146351e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0498313101005311e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 281, + "real_time": 2.4983522041412434e+06, + "cpu_time": 2.8906412775803125e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0492675915168115e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 281, + "real_time": 2.4944472862440706e+06, + "cpu_time": 2.8898351672600727e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0509101613236110e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4946336876402129e+06, + "cpu_time": 2.8899319971532067e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0508329351814370e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time_median", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4944472862440711e+06, + "cpu_time": 2.8898351672600731e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0509101613236110e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1008194551813963e+03, + "cpu_time": 8.9806572963907183e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3061815500725769e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2429959037852177e-03, + "cpu_time": 3.1075669964681941e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2429963948998718e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 271, + "real_time": 2.5823362793677188e+06, + "cpu_time": 3.0482064501843373e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0302855371275315e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 271, + "real_time": 2.5881918961683982e+06, + "cpu_time": 3.0551845202953331e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0256921473873886e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 271, + "real_time": 2.5811377050169501e+06, + "cpu_time": 3.0476618856090852e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0312283183533484e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 271, + "real_time": 2.5850556699758414e+06, + "cpu_time": 3.0466498560884837e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0281497458230743e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 271, + "real_time": 2.5858928679705104e+06, + "cpu_time": 3.0483040295201815e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0274931204380387e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5845228836998837e+06, + "cpu_time": 3.0492013483394841e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0285697738258764e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time_median", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5850556699758414e+06, + "cpu_time": 3.0482064501843369e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0281497458230743e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8223679696827335e+03, + "cpu_time": 3.4085703842656194e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2151773445934302e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0920266899097297e-03, + "cpu_time": 1.1178567745688025e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0919897225992934e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 254, + "real_time": 2.7403224728942858e+06, + "cpu_time": 3.3403207795272712e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8264693676452994e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 254, + "real_time": 2.7477448195673702e+06, + "cpu_time": 3.3457566338582430e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8161331159022886e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 254, + "real_time": 2.7395111030172994e+06, + "cpu_time": 3.3364872125982014e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8276026654723084e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 254, + "real_time": 2.7401964391002390e+06, + "cpu_time": 3.3382656614173008e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8266453639517421e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 254, + "real_time": 2.7336515614572237e+06, + "cpu_time": 3.3305017716532517e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8358070749917996e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7402852792072841e+06, + "cpu_time": 3.3382664118108535e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8265315175926876e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time_median", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7401964391002385e+06, + "cpu_time": 3.3382656614173008e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8266453639517421e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0064852509036191e+03, + "cpu_time": 5.5628014573066412e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9879700943836325e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8269941779024941e-03, + "cpu_time": 1.6663743305882772e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8261890859270487e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 228, + "real_time": 3.0772856369271483e+06, + "cpu_time": 4.0000788289475758e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8149409818652070e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 228, + "real_time": 3.0787200513785998e+06, + "cpu_time": 4.0008418859647582e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8117658150208557e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 228, + "real_time": 3.0772751146159545e+06, + "cpu_time": 3.9901776754384013e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8149642846012676e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 228, + "real_time": 3.0896070601189868e+06, + "cpu_time": 4.0085009429825884e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7877628423053730e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 228, + "real_time": 3.0892288308231193e+06, + "cpu_time": 4.0081765570176374e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7885939010908997e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0824233387727621e+06, + "cpu_time": 4.0015551780701922e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8036055649767208e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time_median", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0787200513785994e+06, + "cpu_time": 4.0008418859647578e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8117658150208557e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4135618841707146e+03, + "cpu_time": 7.4870265773095352e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4146057600900684e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0806882051198762e-03, + "cpu_time": 1.8710291984328608e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0792001337821657e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 164, + "real_time": 3.6231956825160035e+06, + "cpu_time": 5.1077842926828200e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1576255790544026e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 164, + "real_time": 3.6024785552332858e+06, + "cpu_time": 5.0726448658539113e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1642828501801836e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 164, + "real_time": 3.6034082991025615e+06, + "cpu_time": 5.0741224512194097e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1639824443554184e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 164, + "real_time": 3.6058912377395644e+06, + "cpu_time": 5.0766335000001313e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1631809512450230e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 164, + "real_time": 3.6033370983559731e+06, + "cpu_time": 5.0713823536583977e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1640054442626688e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6076621745894775e+06, + "cpu_time": 5.0805134926829347e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1626154538195393e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time_median", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6034082991025611e+06, + "cpu_time": 5.0741224512194097e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1639824443554184e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7763565412459229e+03, + "cpu_time": 1.5369525650566651e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8195560240576882e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4326991044400100e-03, + "cpu_time": 3.0251913852216267e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4251836794312374e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 142, + "real_time": 4.9549815422055172e+06, + "cpu_time": 7.6927324436620036e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6929645304523451e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 142, + "real_time": 4.9475268952705907e+06, + "cpu_time": 7.6810932323945025e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6955153913400218e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 142, + "real_time": 4.9513015095931543e+06, + "cpu_time": 7.7163418521124674e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6942228187370651e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 142, + "real_time": 4.9514243425145531e+06, + "cpu_time": 7.7205867394361161e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6941807891463997e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 142, + "real_time": 4.9507595716752633e+06, + "cpu_time": 7.7198702535207607e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6944082778718781e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9511987722518165e+06, + "cpu_time": 7.7061249042251725e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6942583615095422e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time_median", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9513015095931543e+06, + "cpu_time": 7.7163418521124674e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6942228187370651e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6479408571007502e+03, + "cpu_time": 1.8085960878144499e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.0607373757327278e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3480802910614328e-04, + "cpu_time": 2.3469592178850091e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3479077226804036e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 87, + "real_time": 7.8414595182770966e+06, + "cpu_time": 1.3370762597700136e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1395527147586732e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 87, + "real_time": 7.9317053180487677e+06, + "cpu_time": 1.3523532965517258e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1152091923817544e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 87, + "real_time": 7.8487643683008077e+06, + "cpu_time": 1.3390581011494447e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1375614316769364e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 87, + "real_time": 7.8699911259457301e+06, + "cpu_time": 1.3425131770115212e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1317960505303488e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 87, + "real_time": 7.8324093697485570e+06, + "cpu_time": 1.3347183678160684e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1420249131511621e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8648659400641918e+06, + "cpu_time": 1.3411438404597547e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1332288604997749e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time_median", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8487643683008077e+06, + "cpu_time": 1.3390581011494447e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1375614316769364e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9856994776559812e+04, + "cpu_time": 6.8867482947422599e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0757488454958249e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.0677271653831275e-03, + "cpu_time": 5.1349811161056583e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0428196684147494e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.3717449617152121e+07, + "cpu_time": 2.5018465313725274e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4461130120021710e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.3743159489012232e+07, + "cpu_time": 2.5017632764707275e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4415369716714005e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.3714381183187166e+07, + "cpu_time": 2.4977930098039977e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4466603014604325e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3750870748623913e+07, + "cpu_time": 2.5098813803921591e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4401677983452706e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.3686876999689084e+07, + "cpu_time": 2.4980000647059675e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4515769375849752e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3722547607532904e+07, + "cpu_time": 2.5018568525490757e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4452110042128501e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time_median", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3717449617152119e+07, + "cpu_time": 2.5017632764707275e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4461130120021710e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5456444350155802e+04, + "cpu_time": 4.8936576411519716e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5379408546167724e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8550815109711588e-03, + "cpu_time": 1.9560102474152161e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8558483692402666e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27, + "real_time": 2.5280004643179752e+07, + "cpu_time": 4.8152698148147397e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6546222972354231e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27, + "real_time": 2.5533130944327071e+07, + "cpu_time": 4.8430864148149006e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6283053240249090e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27, + "real_time": 2.5476656785165820e+07, + "cpu_time": 4.8394072592596211e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6341314940143628e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27, + "real_time": 2.5338641540319830e+07, + "cpu_time": 4.8274759962964714e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6484791575433817e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27, + "real_time": 2.5381640151694968e+07, + "cpu_time": 4.8290729962963670e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6439924133712263e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5402014812937491e+07, + "cpu_time": 4.8308624962964199e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6419061372378612e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time_median", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5381640151694968e+07, + "cpu_time": 4.8290729962963670e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6439924133712263e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0256783603460995e+05, + "cpu_time": 1.0957184473298893e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0661672786425214e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.0377834904013659e-03, + "cpu_time": 2.2681631865322637e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0355986293941909e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15, + "real_time": 4.5976659407218300e+07, + "cpu_time": 8.8069820999999136e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9192579393649449e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15, + "real_time": 4.6133630226055779e+07, + "cpu_time": 8.8285410533338413e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9093250919628534e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15, + "real_time": 4.6003699302673340e+07, + "cpu_time": 8.8122344066664487e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9175420680180044e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15, + "real_time": 4.6192106356223427e+07, + "cpu_time": 8.8325462866661534e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9056420801628361e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15, + "real_time": 4.5994377136230469e+07, + "cpu_time": 8.8225377866668478e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9181333970989828e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6060094485680260e+07, + "cpu_time": 8.8205683266666412e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9139801153215241e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time_median", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6003699302673340e+07, + "cpu_time": 8.8225377866668463e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9175420680180044e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6559875311308599e+04, + "cpu_time": 1.0781401190239980e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1029655896260794e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0963889976675656e-03, + "cpu_time": 1.2223023268971552e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0943744802982937e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.6735498160123825e+07, + "cpu_time": 1.6476265887500575e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0948742059962220e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.7685842067003250e+07, + "cpu_time": 1.6712701699999854e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0613317916805863e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.7514041922986507e+07, + "cpu_time": 1.6715732875000811e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0673415385867639e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.7573600932955742e+07, + "cpu_time": 1.6718349525000066e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0652554324619789e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.6679159663617611e+07, + "cpu_time": 1.6514984849999335e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0968857686407881e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7237628549337387e+07, + "cpu_time": 1.6627606967500132e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0771377474732685e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time_median", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7514041922986507e+07, + "cpu_time": 1.6712701699999854e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0673415385867639e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8841569825178303e+05, + "cpu_time": 1.2127382720425513e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7259415242414210e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.5986815136264117e-03, + "cpu_time": 7.2935226001729333e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6089186311488468e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6712774336338043e+08, + "cpu_time": 3.1155459224999046e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2123386649977016e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.7004135251045227e+08, + "cpu_time": 3.1634329599998522e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1572961757464209e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6853203251957893e+08, + "cpu_time": 3.1271664625000995e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1855719294053483e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.7017734795808792e+08, + "cpu_time": 3.1642078400000173e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1547730555315924e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6815091297030449e+08, + "cpu_time": 3.1272779475000334e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1927921324745445e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6880587786436084e+08, + "cpu_time": 3.1395262264999819e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1805543916311221e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time_median", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6853203251957893e+08, + "cpu_time": 3.1272779475000334e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1855719294053483e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2968648260118258e+06, + "cpu_time": 2.2685639929937329e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4447903370751448e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.6825809765574925e-03, + "cpu_time": 7.2258163472097561e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6866798552731354e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.2943345606327057e+08, + "cpu_time": 5.7380436350001669e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2593587695409374e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.2819971442222595e+08, + "cpu_time": 5.7419784149999487e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2716110856167316e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.2859697937965393e+08, + "cpu_time": 5.7258918000002265e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2676557953365164e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.3091129362583160e+08, + "cpu_time": 5.7719948449999952e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2448025941783133e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.2835920155048370e+08, + "cpu_time": 5.7361568950000215e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2700220335836005e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2910012900829315e+08, + "cpu_time": 5.7428131180000710e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2626900556512203e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time_median", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2859697937965393e+08, + "cpu_time": 5.7380436350001669e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2676557953365164e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1185150502662798e+06, + "cpu_time": 1.7361265761720985e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1057350403086627e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.3987074196439888e-03, + "cpu_time": 3.0231291537773440e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3890287506575987e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 311, + "real_time": 2.2557498325692592e+06, + "cpu_time": 2.5811920803858414e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8158041910767776e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 311, + "real_time": 2.2505644680423080e+06, + "cpu_time": 2.5832163954982911e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8199878555636201e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 311, + "real_time": 2.2568153060218049e+06, + "cpu_time": 2.5822285594854308e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8149469250189608e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 311, + "real_time": 2.2545228113339455e+06, + "cpu_time": 2.5797172636653320e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8167924402488074e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 311, + "real_time": 2.2512502568531362e+06, + "cpu_time": 2.5828672829580228e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8194334403877023e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2537805349640907e+06, + "cpu_time": 2.5818443163985834e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8173929704591737e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time_median", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2545228113339455e+06, + "cpu_time": 2.5822285594854308e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8167924402488074e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7560994357804148e+03, + "cpu_time": 1.4165512284491228e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2228889418508370e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2228783561769145e-03, + "cpu_time": 5.4865865437815056e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2231195883239344e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 311, + "real_time": 2.2573338954790323e+06, + "cpu_time": 2.5850292154340879e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6290599349998077e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 311, + "real_time": 2.2500535807126109e+06, + "cpu_time": 2.5787973987137852e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6408021880996828e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 311, + "real_time": 2.2568379437485021e+06, + "cpu_time": 2.5848825401930404e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6298574395614211e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 311, + "real_time": 2.2503545596129545e+06, + "cpu_time": 2.5846542315109842e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6403152405499015e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 311, + "real_time": 2.2569773561497880e+06, + "cpu_time": 2.5850142668812014e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6296332250204128e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2543114671405777e+06, + "cpu_time": 2.5836755305466196e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6339336056462452e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time_median", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2568379437485021e+06, + "cpu_time": 2.5848825401930399e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6298574395614211e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7553906366060100e+03, + "cpu_time": 2.7310975719125131e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0573046400154853e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6658703516996421e-03, + "cpu_time": 1.0570590384214010e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6668726777517107e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2576278175360388e+06, + "cpu_time": 2.5882699512989167e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2571749305788567e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2575379494432504e+06, + "cpu_time": 2.5954883603896466e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2574638242695276e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2632979610995329e+06, + "cpu_time": 2.5932127272726903e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2389938406697847e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2642044668001207e+06, + "cpu_time": 2.5952256883116420e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2360956089600129e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2642616078245468e+06, + "cpu_time": 2.5946983116881852e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2359129984725537e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2613859605406979e+06, + "cpu_time": 2.5933790077922163e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2451282405901477e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time_median", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2632979610995329e+06, + "cpu_time": 2.5946983116881852e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2389938406697847e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4928462326551062e+03, + "cpu_time": 2.9887400165841132e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1196309081602796e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5445599705677690e-03, + "cpu_time": 1.1524501461621971e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5453569225837205e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2825187838684721e+06, + "cpu_time": 2.6183722759739128e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4356070246424848e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2794430162273832e+06, + "cpu_time": 2.6156757175322301e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4375441617414517e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2722611342883343e+06, + "cpu_time": 2.6135705194805381e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4420877735191666e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2724465589498640e+06, + "cpu_time": 2.6146909740260057e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4419701035848625e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2787197769692093e+06, + "cpu_time": 2.6150085292210979e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4380004216043966e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2770778540606527e+06, + "cpu_time": 2.6154636032467568e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4390418970184727e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time_median", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2787197769692093e+06, + "cpu_time": 2.6150085292210979e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4380004216043966e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5426672741669254e+03, + "cpu_time": 1.7953539992730730e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8713231961719321e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9949547469648907e-03, + "cpu_time": 6.8643815079069551e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9953020145702357e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 303, + "real_time": 2.3181448412425937e+06, + "cpu_time": 2.6628765148515333e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8270882316771362e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 303, + "real_time": 2.3198135215604659e+06, + "cpu_time": 2.6659088745874055e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8250546602520011e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 303, + "real_time": 2.3094572706252532e+06, + "cpu_time": 2.6609651518152896e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8377229937775400e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 303, + "real_time": 2.3131579039280270e+06, + "cpu_time": 2.6647430264026928e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8331831514273971e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 303, + "real_time": 2.3179737522140783e+06, + "cpu_time": 2.6634333894390236e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8272968983104933e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3157094579140833e+06, + "cpu_time": 2.6635853914191891e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8300691870889142e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time_median", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3179737522140783e+06, + "cpu_time": 2.6634333894390231e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8272968983104933e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2866575755613994e+03, + "cpu_time": 1.8794918833872841e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2439176755075016e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8511206407658251e-03, + "cpu_time": 7.0562479034541831e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8529291437222909e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 295, + "real_time": 2.3761902272827546e+06, + "cpu_time": 2.7451024508474991e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5160566900354944e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 295, + "real_time": 2.3821590634031314e+06, + "cpu_time": 2.7461995355933742e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5022354306077145e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 295, + "real_time": 2.3761224246315532e+06, + "cpu_time": 2.7494277186440863e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5162140907080710e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 295, + "real_time": 2.3760112381335031e+06, + "cpu_time": 2.7434957118646060e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5164722243891731e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 295, + "real_time": 2.3847082258849326e+06, + "cpu_time": 2.7521600542371036e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4963537500006303e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3790382358671753e+06, + "cpu_time": 2.7472770942373341e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5094664371482171e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time_median", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3761902272827541e+06, + "cpu_time": 2.7461995355933742e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5160566900354944e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1129118227343479e+03, + "cpu_time": 3.4877479427466765e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.5167396777336515e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7288128289519331e-03, + "cpu_time": 1.2695290002098982e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7273432529810743e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 280, + "real_time": 2.4999175752912248e+06, + "cpu_time": 2.8981136357145943e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0486105725684249e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 280, + "real_time": 2.5019872750687813e+06, + "cpu_time": 2.8959274392856774e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0477431384729704e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 280, + "real_time": 2.4948187122520590e+06, + "cpu_time": 2.8920693750003628e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0507537029148065e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 280, + "real_time": 2.4948123800900900e+06, + "cpu_time": 2.8962375499998773e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0507563698659125e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 280, + "real_time": 2.5013278537829006e+06, + "cpu_time": 2.8948223285713731e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0480193534148060e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4985727592970114e+06, + "cpu_time": 2.8954340657143770e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0491766274473841e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time_median", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4999175752912248e+06, + "cpu_time": 2.8959274392856774e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0486105725684249e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5103886761699760e+03, + "cpu_time": 2.2228622202550050e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4745630785083424e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4049575555116695e-03, + "cpu_time": 7.6771294728362896e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4054478911676778e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 271, + "real_time": 2.5851481561497459e+06, + "cpu_time": 3.0527927638376984e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0280771868056539e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 271, + "real_time": 2.5890342344794548e+06, + "cpu_time": 3.0568945276751523e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0250330915589926e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 271, + "real_time": 2.5831372236551172e+06, + "cpu_time": 3.0506437933577574e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0296560136210531e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 271, + "real_time": 2.5881645742226462e+06, + "cpu_time": 3.0495367269371171e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0257135315959173e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 271, + "real_time": 2.5859227673772625e+06, + "cpu_time": 3.0499193837635159e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0274696778038427e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5862813911768454e+06, + "cpu_time": 3.0519574391142484e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0271899002770922e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time_median", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5859227673772625e+06, + "cpu_time": 3.0506437933577569e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0274696778038427e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3676542553787867e+03, + "cpu_time": 3.0336361466854646e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8559873515194008e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.1546660910761296e-04, + "cpu_time": 9.9399687158347101e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.1554686182370479e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 254, + "real_time": 2.7603840823908607e+06, + "cpu_time": 3.3588906771655716e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7986597832131875e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 254, + "real_time": 2.7504105058564679e+06, + "cpu_time": 3.3502263661416047e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8124345357438821e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 254, + "real_time": 2.7472123633850161e+06, + "cpu_time": 3.3451583582680677e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8168727469906348e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 254, + "real_time": 2.7492375738229458e+06, + "cpu_time": 3.3481841850392073e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8140610690908945e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 254, + "real_time": 2.7441830524632079e+06, + "cpu_time": 3.3443567677163016e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8210862029003024e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7502855155836996e+06, + "cpu_time": 3.3493632708661510e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8126228675877810e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time_median", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7492375738229458e+06, + "cpu_time": 3.3481841850392073e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8140610690908945e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1192151586111986e+03, + "cpu_time": 5.8235086372368842e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4670115332388680e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2249381469445374e-03, + "cpu_time": 1.7386912574971049e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2207839136725014e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 228, + "real_time": 3.0778903471256951e+06, + "cpu_time": 3.9959618114032606e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8136020568713152e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 228, + "real_time": 3.0789801579279206e+06, + "cpu_time": 4.0060362324563731e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8111903696428263e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 228, + "real_time": 3.0746804754527514e+06, + "cpu_time": 3.9960361754382523e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8207152474638557e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 228, + "real_time": 3.0730860828618077e+06, + "cpu_time": 3.9989699912279830e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8242540021756554e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 228, + "real_time": 3.0737794535397962e+06, + "cpu_time": 3.9965102938595014e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8227146146900618e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0756833033815939e+06, + "cpu_time": 3.9987029008770743e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8184952581687427e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time_median", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0746804754527514e+06, + "cpu_time": 3.9965102938595014e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8207152474638557e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6036586989000689e+03, + "cpu_time": 4.2802339569146552e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7705994017952762e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.4653016649583108e-04, + "cpu_time": 1.0704055947682008e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4631567278454015e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 194, + "real_time": 3.6017217051679478e+06, + "cpu_time": 5.0729440000001900e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1645275074922593e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 194, + "real_time": 3.6018793135113325e+06, + "cpu_time": 5.0720403350510430e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1644765509678156e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 194, + "real_time": 3.6002941340803332e+06, + "cpu_time": 5.0725289175257906e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1649892602653706e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 194, + "real_time": 3.5949745263481876e+06, + "cpu_time": 5.0626354278350491e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1667131350331478e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 194, + "real_time": 3.5965869349629148e+06, + "cpu_time": 5.0643615412371261e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1661900784954190e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5990913228141433e+06, + "cpu_time": 5.0689020443298398e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1653793064508026e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time_median", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6002941340803327e+06, + "cpu_time": 5.0720403350510430e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1649892602653706e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1368188518638703e+03, + "cpu_time": 4.9806408744000701e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0160111895446822e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.7155856034549833e-04, + "cpu_time": 9.8258771442851205e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7182875474164250e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 141, + "real_time": 4.9565851272262158e+06, + "cpu_time": 7.7001673900711350e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6924168121156428e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 141, + "real_time": 4.9463426727299569e+06, + "cpu_time": 7.6984077801414719e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6959213210697768e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 141, + "real_time": 4.9502066376202917e+06, + "cpu_time": 7.7053161843971647e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6945975418983009e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 141, + "real_time": 4.9509907451768713e+06, + "cpu_time": 7.7099158936175155e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6943291619302595e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 141, + "real_time": 4.9475966787613034e+06, + "cpu_time": 7.6978605106379734e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6954914769043381e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9503443723029289e+06, + "cpu_time": 7.7023335517730536e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6945512627836637e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time_median", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9502066376202917e+06, + "cpu_time": 7.7001673900711359e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6945975418983009e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9669638902159809e+03, + "cpu_time": 5.1599428675972113e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3572409969640616e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.0135109638251835e-04, + "cpu_time": 6.6991942544599467e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.0094419494545262e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 88, + "real_time": 7.7557583678175099e+06, + "cpu_time": 1.3219462409091936e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1631947779106932e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 88, + "real_time": 7.7711056960238656e+06, + "cpu_time": 1.3247394318181127e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1589226367856722e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 88, + "real_time": 7.7760621471415190e+06, + "cpu_time": 1.3286786806818636e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1575465425218220e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 88, + "real_time": 7.7613984097049320e+06, + "cpu_time": 1.3232479863635726e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1616228306256766e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 88, + "real_time": 7.7331201120449062e+06, + "cpu_time": 1.3206481238636445e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1695274038053861e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7594889465465471e+06, + "cpu_time": 1.3238520927272772e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1621628383298502e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time_median", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7613984097049311e+06, + "cpu_time": 1.3232479863635726e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1616228306256766e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6751854722415301e+04, + "cpu_time": 3.0962055033727822e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6744632123057721e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1588863439094032e-03, + "cpu_time": 2.3387850654783245e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1619385595936583e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.3780360055320403e+07, + "cpu_time": 2.5024514686274238e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4349459568035817e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.3757310479002841e+07, + "cpu_time": 2.4977292647057693e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4390255676218553e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.3793574061756041e+07, + "cpu_time": 2.5020050235293731e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4326133205050001e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3747327650586763e+07, + "cpu_time": 2.4977882764705654e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4407967026644506e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.3775067330867637e+07, + "cpu_time": 2.4998168803922098e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4358815237737598e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3770727915506735e+07, + "cpu_time": 2.4999581827450681e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4366526142737293e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time_median", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3775067330867637e+07, + "cpu_time": 2.4998168803922094e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4358815237737598e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8446917371957526e+04, + "cpu_time": 2.2417943095250921e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2644881154631274e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3395746023843150e-03, + "cpu_time": 8.9673272337039637e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3397429310768386e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 28, + "real_time": 2.5235588130142007e+07, + "cpu_time": 4.8012729571427472e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6592946300246325e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 28, + "real_time": 2.3803769078637872e+07, + "cpu_time": 4.5595197071427099e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.8192536979458961e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 28, + "real_time": 2.5249839560793977e+07, + "cpu_time": 4.8062098571429193e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6577936797745643e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 28, + "real_time": 2.5309577118605375e+07, + "cpu_time": 4.8078813821429588e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6515205562509170e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 28, + "real_time": 2.5395057363701720e+07, + "cpu_time": 4.8240135249999739e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6425954877314701e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4998766250376195e+07, + "cpu_time": 4.7597794857142620e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6860916103454962e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time_median", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5249839560793981e+07, + "cpu_time": 4.8062098571429193e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6577936797745643e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7096771396092058e+05, + "cpu_time": 1.1227340593245935e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.4729617733993530e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6840033113667099e-02, + "cpu_time": 2.3587942733362025e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7820949012376202e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15, + "real_time": 4.5510763178269066e+07, + "cpu_time": 8.7397548333334878e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9491425462205305e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15, + "real_time": 4.5511542012294136e+07, + "cpu_time": 8.7394567400004536e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9490920778677082e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15, + "real_time": 4.5583765457073845e+07, + "cpu_time": 8.7567032133332163e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9444195022983041e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15, + "real_time": 4.5560118059317268e+07, + "cpu_time": 8.7394592999999076e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9459477656588688e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15, + "real_time": 4.4457371532917023e+07, + "cpu_time": 8.5305697533332631e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.0190207691568727e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5324712047974274e+07, + "cpu_time": 8.7011887680000648e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9615245322404575e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time_median", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5511542012294129e+07, + "cpu_time": 8.7394592999999076e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9490920778677082e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8588184948912531e+05, + "cpu_time": 9.5667542173915112e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2206033850489568e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0720020658373739e-02, + "cpu_time": 1.0994766890444549e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0874815825390110e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.6213306523859501e+07, + "cpu_time": 1.6519723224999440e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1136197742944775e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.6643931455910206e+07, + "cpu_time": 1.6577495549999809e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0981449189733105e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.7588933296501637e+07, + "cpu_time": 1.6745025437499806e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0647188622711711e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.6383188143372536e+07, + "cpu_time": 1.6543791987500355e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1074965137252212e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.6429128423333168e+07, + "cpu_time": 1.6547396875000685e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1058447643390884e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6651697568595424e+07, + "cpu_time": 1.6586686615000021e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0979649667206540e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time_median", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6429128423333168e+07, + "cpu_time": 1.6547396875000685e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1058447643390884e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4596065932663996e+05, + "cpu_time": 9.0861689842490759e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9386531683454882e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.3006343169958699e-03, + "cpu_time": 5.4779891820178729e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2578279263036550e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6837164014577866e+08, + "cpu_time": 3.1388611899998862e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1886065345397196e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6713074594736099e+08, + "cpu_time": 3.1166410525000286e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2122809537934527e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6848930716514587e+08, + "cpu_time": 3.1356534724997687e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1863797236330404e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.6686786711215973e+08, + "cpu_time": 3.1122797249997801e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2173414887549553e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6814399883151054e+08, + "cpu_time": 3.1311663125001132e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1929234211800442e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6780071184039116e+08, + "cpu_time": 3.1269203504999155e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1995064243802428e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time_median", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6814399883151054e+08, + "cpu_time": 3.1311663125001132e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1929234211800442e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4783390439291159e+05, + "cpu_time": 1.1799270565872099e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4280510540117254e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.4566789746651190e-03, + "cpu_time": 3.7734477515507219e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4633479812072720e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.2075805962085724e+08, + "cpu_time": 5.6147782100003946e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3475131545227122e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.2014526426792145e+08, + "cpu_time": 5.6145920650004649e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3539206848970056e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.2014636695384979e+08, + "cpu_time": 5.6138625450000751e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3539091329272637e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.2007141411304474e+08, + "cpu_time": 5.6066867399999869e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3546945358287120e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.2110342383384705e+08, + "cpu_time": 5.6250724600005245e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3439127218886366e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2044490575790405e+08, + "cpu_time": 5.6149984040002894e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3507900460128665e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time_median", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2014636695384979e+08, + "cpu_time": 5.6145920650004649e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3539091329272637e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6100917751724122e+05, + "cpu_time": 6.5588515603525913e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8169037942645270e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4386534759442654e-03, + "cpu_time": 1.1680949999344386e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4375427072777067e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 312, + "real_time": 2.2411254848031183e+06, + "cpu_time": 2.5714421314102053e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8276531268662233e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 312, + "real_time": 2.2521029138424173e+06, + "cpu_time": 2.5827152532054367e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8187445941409599e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 312, + "real_time": 2.2406085336115211e+06, + "cpu_time": 2.5730151442305581e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8280748013566963e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 312, + "real_time": 2.2405208482478675e+06, + "cpu_time": 2.5776093044873378e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8281463451693186e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 312, + "real_time": 2.2406500640313309e+06, + "cpu_time": 2.5767947243589507e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8280409180140169e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2430015689072511e+06, + "cpu_time": 2.5763153115384979e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8261319571094431e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time_median", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2406500640313304e+06, + "cpu_time": 2.5767947243589507e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8280409180140169e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0932391287386699e+03, + "cpu_time": 4.4017751125301484e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1341148935859464e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2707247285698520e-03, + "cpu_time": 1.7085544975089018e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2638642719608141e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 312, + "real_time": 2.2474798494770834e+06, + "cpu_time": 2.5851462852561101e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6449715008150195e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 312, + "real_time": 2.2535106601515929e+06, + "cpu_time": 2.5859422628207230e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6352168839746811e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 312, + "real_time": 2.2485342014055601e+06, + "cpu_time": 2.5805936794868628e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6432623505922994e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 312, + "real_time": 2.2466798874143609e+06, + "cpu_time": 2.5786228685896834e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6462693443292170e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 312, + "real_time": 2.2408347144030416e+06, + "cpu_time": 2.5790153974360791e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6557805657622316e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2474078625703277e+06, + "cpu_time": 2.5818640987178916e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6451001290946901e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time_median", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2474798494770834e+06, + "cpu_time": 2.5805936794868624e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6449715008150195e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5347046280990189e+03, + "cpu_time": 3.4510435985875847e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3571614908629481e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0177488490730573e-03, + "cpu_time": 1.3366480444502529e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0183702039181565e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 311, + "real_time": 2.2504651481794775e+06, + "cpu_time": 2.5911647363343923e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2802727086237706e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 311, + "real_time": 2.2512142309901030e+06, + "cpu_time": 2.5951106205787421e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2778502260951763e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 311, + "real_time": 2.2562257069615209e+06, + "cpu_time": 2.5937302218648447e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2616848347430974e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 311, + "real_time": 2.2536723625957007e+06, + "cpu_time": 2.5881152926045335e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2699121096420083e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 311, + "real_time": 2.2481515804467383e+06, + "cpu_time": 2.5891110739550665e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2877648208864443e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2519458058347078e+06, + "cpu_time": 2.5914463890675157e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2754969399980996e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time_median", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2512142309901034e+06, + "cpu_time": 2.5911647363343919e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2778502260951763e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0992840723418167e+03, + "cpu_time": 2.9694996801189423e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0010137419253731e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3762693863731920e-03, + "cpu_time": 1.1458850519332803e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3758699236366314e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2655122914927197e+06, + "cpu_time": 2.6103616148864599e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4463836776806694e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2724989883022979e+06, + "cpu_time": 2.6121559773464086e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4419368355573963e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2649649684362621e+06, + "cpu_time": 2.6048657087381696e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4467331926384326e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2653429723769715e+06, + "cpu_time": 2.6162592038834128e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4464917851100180e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2650965245314017e+06, + "cpu_time": 2.6047836440131981e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4466491668287283e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2666831490279306e+06, + "cpu_time": 2.6096852297735298e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4456389315630488e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time_median", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2653429723769715e+06, + "cpu_time": 2.6103616148864594e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4464917851100180e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2580856771003173e+03, + "cpu_time": 4.9252551195090400e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0739758546576970e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4373802877996208e-03, + "cpu_time": 1.8872985382748469e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4346430560051948e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 302, + "real_time": 2.2949274492123169e+06, + "cpu_time": 2.6488557450331720e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8556894041462522e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 302, + "real_time": 2.2953268909974899e+06, + "cpu_time": 2.6435825827812906e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8551924458794512e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 302, + "real_time": 2.2950227902770438e+06, + "cpu_time": 2.6547232417219644e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8555707715690624e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 302, + "real_time": 2.2950795084386077e+06, + "cpu_time": 2.6428929801325868e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8555002020206943e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 302, + "real_time": 2.3034652278740080e+06, + "cpu_time": 2.6577541192052001e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8451048102204997e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2967643733598934e+06, + "cpu_time": 2.6495617337748432e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8534115267671920e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time_median", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2950795084386077e+06, + "cpu_time": 2.6488557450331720e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8555002020206943e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7487990918162855e+03, + "cpu_time": 6.6044925083244261e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6472260665974172e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6322088305176195e-03, + "cpu_time": 2.4926735709286437e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6286560921910025e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 295, + "real_time": 2.3809506250071828e+06, + "cpu_time": 2.7470283118642112e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5050280599415869e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 295, + "real_time": 2.3749937177083250e+06, + "cpu_time": 2.7465644101693663e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5188356509201117e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 295, + "real_time": 2.3803046265206598e+06, + "cpu_time": 2.7462419830506840e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5065220871158257e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 295, + "real_time": 2.3732429283465873e+06, + "cpu_time": 2.7441162644066224e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5229070077253506e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 295, + "real_time": 2.3863331541040186e+06, + "cpu_time": 2.7524932677968508e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4926111123495989e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3791650103373551e+06, + "cpu_time": 2.7472888474575472e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5091807836104959e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time_median", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3803046265206602e+06, + "cpu_time": 2.7465644101693658e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5065220871158257e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2043291391521179e+03, + "cpu_time": 3.1160827790990347e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2046716126967964e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1874603554354419e-03, + "cpu_time": 1.1342392271503539e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1866619739192928e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 282, + "real_time": 2.4921002959607975e+06, + "cpu_time": 2.8884979184396849e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0518998790894718e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 282, + "real_time": 2.4824764097360433e+06, + "cpu_time": 2.8832127624111553e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0559778089809653e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 282, + "real_time": 2.4892639596649306e+06, + "cpu_time": 2.8848273900709972e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0530984429440986e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 282, + "real_time": 2.4957718007292943e+06, + "cpu_time": 2.8914575106387502e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0503524397679242e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 282, + "real_time": 2.4827413759623650e+06, + "cpu_time": 2.8825981347520142e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0558651115982117e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4884707684106864e+06, + "cpu_time": 2.8861187432625210e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0534387364761344e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time_median", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4892639596649306e+06, + "cpu_time": 2.8848273900709972e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0530984429440986e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.8280779264144685e+03, + "cpu_time": 3.7639710970356705e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4669462161749697e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3420319018401378e-03, + "cpu_time": 1.3041636300732411e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3418032114778404e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 271, + "real_time": 2.5821105358826607e+06, + "cpu_time": 3.0485076494459365e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0304630367838958e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 271, + "real_time": 2.5778235644239560e+06, + "cpu_time": 3.0476108413285483e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0338397368834594e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 271, + "real_time": 2.5831207390737161e+06, + "cpu_time": 3.0485627380073555e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0296689661823744e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 271, + "real_time": 2.5816747714246442e+06, + "cpu_time": 3.0481018007378890e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0308057614502794e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 271, + "real_time": 2.5848907253591558e+06, + "cpu_time": 3.0477291439114520e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0282791642078146e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5819240672328267e+06, + "cpu_time": 3.0481024346862370e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0306113331015646e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time_median", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5821105358826607e+06, + "cpu_time": 3.0481018007378886e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0304630367838958e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6050012559003653e+03, + "cpu_time": 4.3505418416851381e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0499666046060360e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0089379811592492e-03, + "cpu_time": 1.4272951565464599e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0095317460259163e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 256, + "real_time": 2.7361910497347708e+06, + "cpu_time": 3.3387034257810200e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8322470212803394e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 256, + "real_time": 2.7434879375505261e+06, + "cpu_time": 3.3519328046871522e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8220543478540033e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 256, + "real_time": 2.7360398098608130e+06, + "cpu_time": 3.3355474375005299e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8324588561207479e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 256, + "real_time": 2.7336779221514007e+06, + "cpu_time": 3.3409960078127910e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8357700865315259e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 256, + "real_time": 2.7370654352125712e+06, + "cpu_time": 3.3365578828128404e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8310227680711752e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7372924309020163e+06, + "cpu_time": 3.3407475117188664e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8307106159715587e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time_median", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7361910497347708e+06, + "cpu_time": 3.3387034257810200e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8322470212803394e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6838364217601761e+03, + "cpu_time": 6.5944025254855715e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1488723009995109e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3457957141050677e-03, + "cpu_time": 1.9739302363777408e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3441036969830296e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 227, + "real_time": 3.0720440798032912e+06, + "cpu_time": 3.9999133127751434e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8265687129537690e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 227, + "real_time": 3.0723704837492383e+06, + "cpu_time": 3.9981316475767633e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8258434687239563e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 227, + "real_time": 3.0671185885642199e+06, + "cpu_time": 3.9906367621143637e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8375315118862724e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 227, + "real_time": 3.0663348149737990e+06, + "cpu_time": 4.0032843964766148e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8392792259964585e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 227, + "real_time": 3.0680708453830513e+06, + "cpu_time": 3.9984487180617088e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8354093033929563e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0691877624947205e+06, + "cpu_time": 3.9980829674009187e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8329264445906830e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time_median", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0680708453830513e+06, + "cpu_time": 3.9984487180617088e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8354093033929563e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8265065580499509e+03, + "cpu_time": 4.6364626896783320e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2912312569162482e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.2092982794656018e-04, + "cpu_time": 1.1596714544151675e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.2072281297521212e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 195, + "real_time": 3.5889030792392218e+06, + "cpu_time": 5.0652147794876210e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1686868960777597e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 195, + "real_time": 3.5891673479897855e+06, + "cpu_time": 5.0647088410245581e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1686008461960232e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 195, + "real_time": 3.5898384327689805e+06, + "cpu_time": 5.0687555487177009e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1683823878292961e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 195, + "real_time": 3.5954237855875338e+06, + "cpu_time": 5.0799495230771406e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1665673506453154e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 195, + "real_time": 3.5898713227839042e+06, + "cpu_time": 5.0714617128202617e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1683716832355332e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5906407936738855e+06, + "cpu_time": 5.0700180810254561e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1681218327967854e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time_median", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5898384327689805e+06, + "cpu_time": 5.0687555487176999e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1683823878292961e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7066387710035588e+03, + "cpu_time": 6.1961951401327387e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7969870914989978e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5380382681893661e-04, + "cpu_time": 1.2221248605250542e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5308814924182513e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 145, + "real_time": 4.8355869306572555e+06, + "cpu_time": 7.5353626689649960e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7347652147078278e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 145, + "real_time": 4.8690844776815381e+06, + "cpu_time": 7.5788123517229212e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7228306550134692e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 145, + "real_time": 4.8737322462016139e+06, + "cpu_time": 7.5839925448266948e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7211877009734657e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 145, + "real_time": 4.8731518395501990e+06, + "cpu_time": 7.5946835379313435e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7213926994678426e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 145, + "real_time": 4.9455726487112455e+06, + "cpu_time": 7.6861900137922438e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6961853754561601e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8794256285603708e+06, + "cpu_time": 7.5958082234476404e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7192723291237531e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time_median", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8731518395501990e+06, + "cpu_time": 7.5839925448266957e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7213926994678426e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0236718744750215e+04, + "cpu_time": 5.5344981518201515e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4087789360174293e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.2461998209862435e-03, + "cpu_time": 7.2862531399036741e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1940418173043570e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 88, + "real_time": 7.7479400894265957e+06, + "cpu_time": 1.3257060852272285e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1653776108691669e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 88, + "real_time": 7.7318617841228843e+06, + "cpu_time": 1.3195417545455772e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1698804852476082e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 88, + "real_time": 7.7215527046725834e+06, + "cpu_time": 1.3191665272725832e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1727775023600521e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 88, + "real_time": 7.7315345882776789e+06, + "cpu_time": 1.3202483431818090e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1699723138323812e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 88, + "real_time": 7.7531655498949643e+06, + "cpu_time": 1.3243108897726912e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1639181947078753e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7372109432789413e+06, + "cpu_time": 1.3217947199999779e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1683852214034166e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time_median", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7318617841228843e+06, + "cpu_time": 1.3202483431818087e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1698804852476082e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2996913050112231e+04, + "cpu_time": 3.0001841560126060e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6416636641183654e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6797930346467056e-03, + "cpu_time": 2.2697807084693573e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6794357516241589e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.3737269476348279e+07, + "cpu_time": 2.4958427039216664e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4425838087963052e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.3709225454459004e+07, + "cpu_time": 2.4910494843135323e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4475804349024134e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.3778170077678036e+07, + "cpu_time": 2.5027808529410534e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4353329804196143e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3758756676871402e+07, + "cpu_time": 2.4957122647060268e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4387691989934897e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.3745705651886323e+07, + "cpu_time": 2.4977176137255415e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4410847176401834e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3745825467448607e+07, + "cpu_time": 2.4966205839215644e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4410702281504016e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time_median", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3745705651886323e+07, + "cpu_time": 2.4958427039216664e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4410847176401834e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5623086652473237e+04, + "cpu_time": 4.2314205864787669e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5520510480090184e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8640631450727412e-03, + "cpu_time": 1.6948592884835820e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8647767669749127e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 28, + "real_time": 2.5315405601369482e+07, + "cpu_time": 4.8155784607144728e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6509100844258099e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 28, + "real_time": 2.5268279481679201e+07, + "cpu_time": 4.8100264071430437e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6558541134016414e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 28, + "real_time": 2.5361970865300722e+07, + "cpu_time": 4.8156133428571068e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6460429418683615e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 28, + "real_time": 2.5338599657905955e+07, + "cpu_time": 4.8158008392858520e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6484835352399282e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 28, + "real_time": 2.5359771014856439e+07, + "cpu_time": 4.8149506749999605e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6462724746483636e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5328805324222364e+07, + "cpu_time": 4.8143939450000875e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6495126299168210e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time_median", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5338599657905951e+07, + "cpu_time": 4.8155784607144728e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6484835352399282e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8722365216854698e+04, + "cpu_time": 2.4624528090321612e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0547400986006488e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5287876676845810e-03, + "cpu_time": 5.1147721544255896e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5303720589276617e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15, + "real_time": 4.5137899617354073e+07, + "cpu_time": 8.6562828600002229e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9735040650495305e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15, + "real_time": 4.5698658128579460e+07, + "cpu_time": 8.7483092866659716e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9370168292985754e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15, + "real_time": 4.5572427163521446e+07, + "cpu_time": 8.7347057733328864e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9451520657086020e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15, + "real_time": 4.5548406243324280e+07, + "cpu_time": 8.7433825133333206e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9467052542518187e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15, + "real_time": 4.5594539741675057e+07, + "cpu_time": 8.7356615733324364e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9437237169283266e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5510386178890862e+07, + "cpu_time": 8.7236684013329685e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9492203862473712e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time_median", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5572427163521446e+07, + "cpu_time": 8.7356615733324364e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9451520657086020e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1596816708011119e+05, + "cpu_time": 3.8086383320828795e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4068709543892076e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.7454698852958535e-03, + "cpu_time": 4.3658678400716414e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7703147616558073e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.4820338524878025e+07, + "cpu_time": 1.6229351662499881e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1647534149049311e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.6297065950930119e+07, + "cpu_time": 1.6518500062500152e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1105977131671739e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.6418905295431614e+07, + "cpu_time": 1.6522167949997878e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1062121775591431e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.6384426802396774e+07, + "cpu_time": 1.6525678275002065e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1074519555942941e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.4718359634280205e+07, + "cpu_time": 1.6207006262499136e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1685629556427460e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.5727819241583347e+07, + "cpu_time": 1.6400540842499822e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1315156433736577e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time_median", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6297065950930119e+07, + "cpu_time": 1.6518500062500152e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1105977131671739e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7682740445540531e+05, + "cpu_time": 1.6667954049596041e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2148635502211910e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0228038135257840e-02, + "cpu_time": 1.0163051456451518e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0266158360166262e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6411680728197098e+08, + "cpu_time": 3.0592655249995458e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2712731918894567e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6760784015059471e+08, + "cpu_time": 3.1280565324999541e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2031372250702858e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6401332244277000e+08, + "cpu_time": 3.0513230075001729e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2733372143432622e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.6760527342557907e+08, + "cpu_time": 3.1301496300000054e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2031862782550459e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6397401317954063e+08, + "cpu_time": 3.0546471625001460e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2741219269431558e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6546345129609108e+08, + "cpu_time": 3.0846883714999646e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2450111673002415e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time_median", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6411680728197098e+08, + "cpu_time": 3.0592655249995458e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2712731918894567e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9570740461677960e+06, + "cpu_time": 4.0649619586503282e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8217285307554744e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1827832858784509e-02, + "cpu_time": 1.3177869104079689e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1777243077825971e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.1779828667640686e+08, + "cpu_time": 5.5897992749999046e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3786897822181177e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.1768591701984406e+08, + "cpu_time": 5.5706606900002956e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3798848689063210e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.1924095749855042e+08, + "cpu_time": 5.5791950900004399e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3634212615243006e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.1777060031890869e+08, + "cpu_time": 5.5677715700005591e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3789841568805060e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.1719118356704712e+08, + "cpu_time": 5.5675439399999499e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3851565857694621e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1793738901615143e+08, + "cpu_time": 5.5749941130002296e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3772273310597420e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time_median", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1777060031890869e+08, + "cpu_time": 5.5706606900002956e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3789841568805060e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6917479417134426e+05, + "cpu_time": 9.5300811855434626e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1512015067227976e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_coarse/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4192649897249725e-03, + "cpu_time": 1.7094334078883481e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4135779761574497e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + } + ] +} diff --git a/results/implicit_managed_GPUWrGPU_fine.json b/results/implicit_managed_GPUWrGPU_fine.json new file mode 100644 index 0000000..c07aa39 --- /dev/null +++ b/results/implicit_managed_GPUWrGPU_fine.json @@ -0,0 +1,25272 @@ +{ + "context": { + "date": "2023-10-09T17:38:29-04:00", + "host_name": "frontier08444", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1804, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.71,11.7,26.82], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52237, + "real_time": 1.3400284570844202e+04, + "cpu_time": 2.5679946551295056e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0566515049328965e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52237, + "real_time": 1.3399247187575495e+04, + "cpu_time": 2.5717080689932420e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0568881539837790e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52237, + "real_time": 1.3397628338803852e+04, + "cpu_time": 2.5488320500794453e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0572575208230424e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52237, + "real_time": 1.3400425779468309e+04, + "cpu_time": 2.6014288033386289e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0566192950941575e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52237, + "real_time": 1.3398446076847365e+04, + "cpu_time": 2.5689979152707841e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0570709293504745e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3399206390707846e+04, + "cpu_time": 2.5717922985623209e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0568974808368701e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3399247187575495e+04, + "cpu_time": 2.5689979152707841e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0568881539837790e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1957602766579982e+00, + "cpu_time": 1.8892678436342032e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7280561577797478e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.9241126809363921e-05, + "cpu_time": 7.3461136215795444e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9242644703704712e-05, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52260, + "real_time": 1.3397411057834412e+04, + "cpu_time": 2.5618317374665148e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1146142076528728e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52260, + "real_time": 1.3395386904925819e+04, + "cpu_time": 2.5654872139303487e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1155381760474551e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52260, + "real_time": 1.3395920566535411e+04, + "cpu_time": 2.5540653195560662e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1152945475539637e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52260, + "real_time": 1.3399812355815231e+04, + "cpu_time": 2.5387169230769236e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1135184452376664e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52260, + "real_time": 1.3396265476355247e+04, + "cpu_time": 2.5885375296593968e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1151370988124192e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3396959272293223e+04, + "cpu_time": 2.5617277447378503e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1148204950608754e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3396265476355245e+04, + "cpu_time": 2.5618317374665152e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1151370988124192e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7590336385689216e+00, + "cpu_time": 1.8173604924559288e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.0279479320683196e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3130096186877627e-04, + "cpu_time": 7.0942764944051662e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3128673096050384e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52260, + "real_time": 1.3393965895471767e+04, + "cpu_time": 2.5673377937236892e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2232373986810811e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52260, + "real_time": 1.3393881163649170e+04, + "cpu_time": 2.5588281687715218e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2232451370754263e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52260, + "real_time": 1.3398118764255429e+04, + "cpu_time": 2.5600693628013763e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2228582451224828e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52260, + "real_time": 1.3399794253076754e+04, + "cpu_time": 2.5929857462686530e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2227053408852181e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52260, + "real_time": 1.3398538894421064e+04, + "cpu_time": 2.5854447818599267e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2228199006700673e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3396859794174839e+04, + "cpu_time": 2.5729331706850335e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2229732044868550e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3398118764255432e+04, + "cpu_time": 2.5673377937236895e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2228582451224828e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7505600257369878e+00, + "cpu_time": 1.5446545546691760e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5110482464500758e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0531378755885569e-04, + "cpu_time": 6.0034771686585144e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0532324316162607e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52257, + "real_time": 1.3396556343471111e+04, + "cpu_time": 2.5993877337007412e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4460017305842686e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52257, + "real_time": 1.3396470301516483e+04, + "cpu_time": 2.5722521021107212e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4460174406008019e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52257, + "real_time": 1.3396694067148455e+04, + "cpu_time": 2.5622723903017755e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4459765846526351e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52257, + "real_time": 1.3395108655901766e+04, + "cpu_time": 2.5828989475094269e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4462660842667155e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52257, + "real_time": 1.3397598307154392e+04, + "cpu_time": 2.6021560977476693e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4458114991029181e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3396485535038442e+04, + "cpu_time": 2.5837934542740666e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4460146678414683e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3396556343471113e+04, + "cpu_time": 2.5828989475094269e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4460017305842686e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9201868633175463e-01, + "cpu_time": 1.7157559360148937e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6287505689945284e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.6586022430934161e-05, + "cpu_time": 6.6404531413945629e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6587931397478083e-05, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52254, + "real_time": 1.3395963961299001e+04, + "cpu_time": 2.5978959314119438e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.8922197901796227e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52254, + "real_time": 1.3393072908105178e+04, + "cpu_time": 2.6228325927201658e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.8932758336840782e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52254, + "real_time": 1.3395101923269811e+04, + "cpu_time": 2.6269395586940638e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.8925346276127729e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52254, + "real_time": 1.3396511706824154e+04, + "cpu_time": 2.5978742488613334e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.8920197611305113e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52254, + "real_time": 1.3395873094538645e+04, + "cpu_time": 2.6182645481685573e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.8922529750388823e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3395304718807360e+04, + "cpu_time": 2.6127613759712134e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.8924605975291739e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3395873094538643e+04, + "cpu_time": 2.6182645481685573e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.8922529750388823e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3450442829755485e+00, + "cpu_time": 1.3922492340533123e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9130500506304635e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0041162267007416e-04, + "cpu_time": 5.3286505490222451e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0042084044809044e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52237, + "real_time": 1.3402223187851165e+04, + "cpu_time": 2.6188396672856401e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.7798699635754471e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52237, + "real_time": 1.3401130935258263e+04, + "cpu_time": 2.5993205926833511e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.7806670670719795e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52237, + "real_time": 1.3402527706219347e+04, + "cpu_time": 2.6056862434672737e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.7796477554884663e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52237, + "real_time": 1.3402107254185279e+04, + "cpu_time": 2.6068673737006346e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.7799545634189835e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52237, + "real_time": 1.3404927763791016e+04, + "cpu_time": 2.6380477611654638e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.7778967786792355e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3402583369461014e+04, + "cpu_time": 2.6137523276604727e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.7796072256468239e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3402223187851163e+04, + "cpu_time": 2.6068673737006342e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 9.7798699635754471e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4111400314886084e+00, + "cpu_time": 1.5304316330235744e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0295885102700010e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0528865910314106e-04, + "cpu_time": 5.8553047158582125e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0527912691318786e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51062, + "real_time": 1.3710655457486910e+04, + "cpu_time": 2.6662151071246786e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9119727777628044e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51062, + "real_time": 1.3709835418736842e+04, + "cpu_time": 2.6622715385217900e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9120871403148666e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51062, + "real_time": 1.3707747920392687e+04, + "cpu_time": 2.6443628079589536e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9123783244512009e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51062, + "real_time": 1.3707359312963476e+04, + "cpu_time": 2.6339304355489297e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9124325409058350e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51062, + "real_time": 1.3712281519485245e+04, + "cpu_time": 2.6799538423876929e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9117460477127136e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3709575925813033e+04, + "cpu_time": 2.6573467463084089e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9121233662294842e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3709835418736842e+04, + "cpu_time": 2.6622715385217900e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.9120871403148666e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0498742930955958e+00, + "cpu_time": 1.8240671997601333e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8589820801117308e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4952134947048188e-04, + "cpu_time": 6.8642423210073383e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4951870421150480e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50898, + "real_time": 1.3757761828257515e+04, + "cpu_time": 2.6994015481944276e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8108524231255974e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50898, + "real_time": 1.3753329564147476e+04, + "cpu_time": 2.6833001768242455e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8120805406039795e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50898, + "real_time": 1.3757038854393855e+04, + "cpu_time": 2.6643768399544162e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8110526949085983e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50898, + "real_time": 1.3755440043618824e+04, + "cpu_time": 2.6516704133757645e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8114956579903694e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50898, + "real_time": 1.3758258232263361e+04, + "cpu_time": 2.6889199909623207e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8107149258947273e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3756365704536205e+04, + "cpu_time": 2.6775337938622350e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8112392485046547e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3757038854393855e+04, + "cpu_time": 2.6833001768242459e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.8110526949085983e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0036809880804540e+00, + "cpu_time": 1.9255617509692252e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5517306475296514e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4565482127446884e-04, + "cpu_time": 7.1915497588983923e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4566733509862654e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50622, + "real_time": 1.3829097961651778e+04, + "cpu_time": 2.6977685038125670e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.5823889808844467e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50622, + "real_time": 1.3830643712235093e+04, + "cpu_time": 2.6953179724230624e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.5815415523457626e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50622, + "real_time": 1.3835577941691370e+04, + "cpu_time": 2.6786927738927840e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.5788377212655396e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50622, + "real_time": 1.3832129311247765e+04, + "cpu_time": 2.7059047726285025e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.5807272792579941e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50622, + "real_time": 1.3835013719615596e+04, + "cpu_time": 2.6815579293587718e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.5791468028203339e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3832492529288322e+04, + "cpu_time": 2.6918483904231380e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.5805284673148148e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3832129311247767e+04, + "cpu_time": 2.6953179724230624e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 7.5807272792579941e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7816052694056337e+00, + "cpu_time": 1.1441497473109412e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5243848965801256e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0109212157649702e-04, + "cpu_time": 4.2504241746359637e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0109216701089645e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50043, + "real_time": 1.3985155716072379e+04, + "cpu_time": 2.7057248925923643e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4995557021863242e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50043, + "real_time": 1.3985335592162608e+04, + "cpu_time": 2.7376555742061846e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4995364152543079e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50043, + "real_time": 1.3988992424751432e+04, + "cpu_time": 2.7543972563595256e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4991444246473413e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50043, + "real_time": 1.3985526739658861e+04, + "cpu_time": 2.7184656895070108e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4995159203071634e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50043, + "real_time": 1.3990220615680120e+04, + "cpu_time": 2.7349236176887815e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4990128158876422e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3987046217665082e+04, + "cpu_time": 2.7302334060707733e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4993530556565558e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3985526739658861e+04, + "cpu_time": 2.7349236176887818e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.4995159203071634e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3808388136544325e+00, + "cpu_time": 1.8709120180245927e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5519661665268213e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7021741235454904e-04, + "cpu_time": 6.8525717027143242e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7020448632155778e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49877, + "real_time": 1.4037276396124815e+04, + "cpu_time": 2.7440417246426161e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.9879756454449353e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49877, + "real_time": 1.4031587798575954e+04, + "cpu_time": 2.7487853459510490e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.9891870116264917e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49877, + "real_time": 1.4038710007311560e+04, + "cpu_time": 2.7423082382661352e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.9876705180287549e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49877, + "real_time": 1.4034857260184528e+04, + "cpu_time": 2.7334771959019148e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.9884906716499469e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49877, + "real_time": 1.4033496673269579e+04, + "cpu_time": 2.7481398500310854e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.9887804142136121e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4035185627093288e+04, + "cpu_time": 2.7433504709585606e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.9884208521927484e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4034857260184528e+04, + "cpu_time": 2.7440417246426157e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.9884906716499469e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8590790052041517e+00, + "cpu_time": 6.1547067164782071e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0876314954501636e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0370795806825905e-04, + "cpu_time": 2.2434999762635786e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0370730216874824e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 39809, + "real_time": 1.7552855703223860e+04, + "cpu_time": 3.1275718505865687e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7790559791699872e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 39809, + "real_time": 1.7570999924582506e+04, + "cpu_time": 3.1176743726292782e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7741210153122900e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 39809, + "real_time": 1.7616855476839009e+04, + "cpu_time": 3.0847599361953580e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7616942825174194e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 39809, + "real_time": 1.7625162710698867e+04, + "cpu_time": 3.0803920018086326e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7594499623586041e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 39809, + "real_time": 1.7566895050687916e+04, + "cpu_time": 3.1203664975256881e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7752365889335150e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7586553773206433e+04, + "cpu_time": 3.1061529317491055e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7699115656583636e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7570999924582509e+04, + "cpu_time": 3.1176743726292782e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.7741210153122900e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2298571160529164e+01, + "cpu_time": 2.1879356174030622e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7559197624105299e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8365491941767962e-03, + "cpu_time": 7.0438760276076099e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8356566242129062e-03, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27719, + "real_time": 2.5253497997514063e+04, + "cpu_time": 3.9160754969515401e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.6435216228862781e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27719, + "real_time": 2.5248854938897715e+04, + "cpu_time": 3.8985223601140235e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.6447433123604614e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27719, + "real_time": 2.5252289080514605e+04, + "cpu_time": 3.8945594141202993e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.6438396719233594e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27719, + "real_time": 2.5251175331394934e+04, + "cpu_time": 3.9112014286229583e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.6441327105834912e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27719, + "real_time": 2.5249610726730913e+04, + "cpu_time": 3.9339168476496190e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.6445444175654260e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5251085615010445e+04, + "cpu_time": 3.9108551094916882e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.6441563470638037e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5251175331394927e+04, + "cpu_time": 3.9112014286229576e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.6441327105834912e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8991562613765023e+00, + "cpu_time": 1.5627400166653746e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9971080812972620e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5210875695877146e-05, + "cpu_time": 3.9959036397758316e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5210573325921087e-05, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15889, + "real_time": 4.4022590313416040e+04, + "cpu_time": 5.8011982818302778e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6220939661004382e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15889, + "real_time": 4.4031962651223403e+04, + "cpu_time": 5.8387643464031433e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6204715801074353e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15889, + "real_time": 4.4023425848322346e+04, + "cpu_time": 5.8184267732393404e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6219493039019592e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15889, + "real_time": 4.4038367593964264e+04, + "cpu_time": 5.8001235131222973e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6193632582781848e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15889, + "real_time": 4.4061376946045115e+04, + "cpu_time": 5.8609971049153683e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6153843401418701e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4035544670594230e+04, + "cpu_time": 5.8239020039020848e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6198524897059778e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4031962651223403e+04, + "cpu_time": 5.8184267732393404e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6204715801074353e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5833409576069608e+01, + "cpu_time": 2.6018212439909490e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7389700156401622e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.5955975325184834e-04, + "cpu_time": 4.4674880213432452e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5945184232114301e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8470, + "real_time": 8.2636942916275962e+04, + "cpu_time": 9.8081457615111853e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 8.1209277148589209e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8470, + "real_time": 8.2625491770219116e+04, + "cpu_time": 9.7937581109799328e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 8.1220532020104944e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8470, + "real_time": 8.2650333254507641e+04, + "cpu_time": 9.7902418299881450e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 8.1196120278607556e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8470, + "real_time": 8.2662371134817324e+04, + "cpu_time": 9.7896502833530089e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 8.1184295924138818e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8470, + "real_time": 8.2630011474825282e+04, + "cpu_time": 9.8184193388428495e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 8.1216089411346521e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.2641030110129068e+04, + "cpu_time": 9.8000430649350252e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 8.1205262956557422e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.2636942916275948e+04, + "cpu_time": 9.7937581109799328e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 8.1209277148589209e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5178465943753807e+01, + "cpu_time": 1.2717270699910571e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4913711493611777e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8366743400374710e-04, + "cpu_time": 1.2976749811859003e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8365449418703567e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4397, + "real_time": 1.5895635124035110e+05, + "cpu_time": 1.7799103616101708e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 8.4436845053806714e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4397, + "real_time": 1.5898494433730457e+05, + "cpu_time": 1.7822071389583687e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 8.4421659270604822e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4397, + "real_time": 1.5900153878278570e+05, + "cpu_time": 1.7813265703888953e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 8.4412848471458362e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4397, + "real_time": 1.5897933733907959e+05, + "cpu_time": 1.7812361314532885e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 8.4424636714728088e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4397, + "real_time": 1.5899501798253178e+05, + "cpu_time": 1.7821807323174982e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 8.4416310462473755e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5898343793641057e+05, + "cpu_time": 1.7813721869456448e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 8.4422459994614343e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5898494433730457e+05, + "cpu_time": 1.7813265703888953e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 8.4421659270604822e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7424919787850758e+01, + "cpu_time": 9.3653827470740580e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.2534766571850598e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0960210707495388e-04, + "cpu_time": 5.2573980977731671e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0960918051636232e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2212, + "real_time": 3.1654137995034695e+05, + "cpu_time": 3.5351854339964077e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.4802642877878113e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2212, + "real_time": 3.1590266558985860e+05, + "cpu_time": 3.5372118851717899e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.4974102861326904e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2212, + "real_time": 3.1630512760671286e+05, + "cpu_time": 3.5323245569620479e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.4865983056008813e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2212, + "real_time": 3.1637384017155156e+05, + "cpu_time": 3.5341999547920545e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.4847551192741687e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2212, + "real_time": 3.1636828172149451e+05, + "cpu_time": 3.5351384222423198e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.4849041926494153e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1629825900799292e+05, + "cpu_time": 3.5348120506329241e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.4867864382889941e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1636828172149451e+05, + "cpu_time": 3.5351384222423198e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.4849041926494153e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3783650651341685e+02, + "cpu_time": 1.7725282393962348e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.3857084502468312e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5193745061811004e-04, + "cpu_time": 5.0144907678439531e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5242949692206967e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1106, + "real_time": 6.2867504669602611e+05, + "cpu_time": 7.2002871518987685e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.5397203980259961e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1106, + "real_time": 6.3335804763851292e+05, + "cpu_time": 7.2574059132008080e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.4765783588245715e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1106, + "real_time": 6.2914984784919303e+05, + "cpu_time": 7.2049390958407568e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.5332757185802856e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1106, + "real_time": 6.3206357041468157e+05, + "cpu_time": 7.2333955244124145e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.4939385392480701e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1106, + "real_time": 6.2907317178163875e+05, + "cpu_time": 7.1996428300180647e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.5343158170216235e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3046393687601050e+05, + "cpu_time": 7.2191341030741634e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.5155657663401099e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2914984784919303e+05, + "cpu_time": 7.2049390958407579e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.5332757185802856e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1092587654023764e+03, + "cpu_time": 2.5518302810253963e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8444718038896575e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.3455660855938722e-03, + "cpu_time": 3.5348149024392503e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3403203990663068e-03, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 532, + "real_time": 1.3090368512146091e+06, + "cpu_time": 1.6400401109022477e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 8.2025332060263452e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 532, + "real_time": 1.3117219940269073e+06, + "cpu_time": 1.6428548834586511e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 8.1857423210819055e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 532, + "real_time": 1.3120925229890062e+06, + "cpu_time": 1.6425988966165413e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 8.1834307046729260e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 532, + "real_time": 1.3107141918376705e+06, + "cpu_time": 1.6418816503759825e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 8.1920363011754199e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 532, + "real_time": 1.3121057866767917e+06, + "cpu_time": 1.6429404398496288e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 8.1833479808018909e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3111342693489972e+06, + "cpu_time": 1.6420631962406104e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 8.1894181027516980e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3117219940269070e+06, + "cpu_time": 1.6425988966165413e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 8.1857423210819055e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3022235021317936e+03, + "cpu_time": 1.2051071611861901e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1404481454810059e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.9320377216467097e-04, + "cpu_time": 7.3389816174261699e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9402033738462585e-04, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 303, + "real_time": 2.2845314377642209e+06, + "cpu_time": 2.6024208943893882e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.7929278329426670e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 303, + "real_time": 2.2729780162275820e+06, + "cpu_time": 2.5721711089108428e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8020411859495468e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 303, + "real_time": 2.2678337107556979e+06, + "cpu_time": 2.5675092937293844e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8061288976232354e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 303, + "real_time": 2.2745416138517875e+06, + "cpu_time": 2.5799104026402584e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8008024012643548e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 303, + "real_time": 2.2680206414889973e+06, + "cpu_time": 2.5677612871287740e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8059800361035957e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2735810840176572e+06, + "cpu_time": 2.5779545973597295e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8015760707766798e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time_median", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2729780162275820e+06, + "cpu_time": 2.5721711089108423e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8020411859495468e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8034435402933350e+03, + "cpu_time": 1.4567785904748362e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3788952669300543e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9923909853573085e-03, + "cpu_time": 5.6509086388364990e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9856609188926180e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2694035667575314e+06, + "cpu_time": 2.5822853863636903e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6097590221489472e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2691381012706403e+06, + "cpu_time": 2.5715724415584467e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6101813262986327e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2833974443847775e+06, + "cpu_time": 2.5854041233767010e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.5876364932199507e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2693791437301445e+06, + "cpu_time": 2.5713832824674700e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6097978703263006e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2765345132135628e+06, + "cpu_time": 2.5785287662338465e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.5984519243840273e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2735705538713313e+06, + "cpu_time": 2.5778348000000310e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6031653272755723e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time_median", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2694035667575309e+06, + "cpu_time": 2.5785287662338465e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6097590221489472e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3231947774634127e+03, + "cpu_time": 6.2933134400791023e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0000656673751479e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7811737650703505e-03, + "cpu_time": 2.4413175895053579e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7755197903486660e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 305, + "real_time": 2.2751504196556378e+06, + "cpu_time": 2.5796950918033011e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2012821035717940e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 305, + "real_time": 2.2835014082613541e+06, + "cpu_time": 2.5944730655737659e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1749463086491777e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 305, + "real_time": 2.2749563060761965e+06, + "cpu_time": 2.5794719213114954e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2018965622503879e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 305, + "real_time": 2.2762866576248016e+06, + "cpu_time": 2.5887492163934386e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1976874903338999e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 305, + "real_time": 2.2813489729324812e+06, + "cpu_time": 2.5880349901638799e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1817158156824000e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2782487529100943e+06, + "cpu_time": 2.5860848570491769e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1915056560975313e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time_median", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2762866576248016e+06, + "cpu_time": 2.5880349901638804e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1976874903338999e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9208240664139112e+03, + "cpu_time": 6.4386312810467389e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2367476437717600e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7209815483957545e-03, + "cpu_time": 2.4897215818329593e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7197339512945342e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 306, + "real_time": 2.2953041133720404e+06, + "cpu_time": 2.6036699836601233e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4276103897997377e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 306, + "real_time": 2.2975880587230027e+06, + "cpu_time": 2.6065813235294251e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4261912563304504e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 306, + "real_time": 2.2897244631951931e+06, + "cpu_time": 2.6041484901961302e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4310892217255669e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 306, + "real_time": 2.2901401480696364e+06, + "cpu_time": 2.6036760130719338e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4308294637609934e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 306, + "real_time": 2.2964816703308932e+06, + "cpu_time": 2.6052283333333796e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4268783602038745e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2938476907381536e+06, + "cpu_time": 2.6046608287581988e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4285197383641247e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time_median", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2953041133720400e+06, + "cpu_time": 2.6041484901961302e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4276103897997377e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6672962130215251e+03, + "cpu_time": 1.2474462267144354e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2847287481743595e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5987531464399017e-03, + "cpu_time": 4.7892847043320009e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5993679938862632e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 301, + "real_time": 2.3269730502510588e+06, + "cpu_time": 2.6446252225913098e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8163626558945011e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 301, + "real_time": 2.3207425463177338e+06, + "cpu_time": 2.6448426877076058e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8239237525069032e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 301, + "real_time": 2.3287952447501924e+06, + "cpu_time": 2.6467400033223224e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8141589582741518e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 301, + "real_time": 2.3202375652720267e+06, + "cpu_time": 2.6379666677740091e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8245383568003953e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 301, + "real_time": 2.3206952430208856e+06, + "cpu_time": 2.6498078936876906e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8239813132331308e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3234887299223798e+06, + "cpu_time": 2.6447964950165884e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8205930073418166e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time_median", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3207425463177343e+06, + "cpu_time": 2.6448426877076062e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8239237525069032e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0686223963459206e+03, + "cpu_time": 4.3465316444697010e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9354109072535692e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7510833359978639e-03, + "cpu_time": 1.6434276333394944e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7497777575166008e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 294, + "real_time": 2.3858862791248425e+06, + "cpu_time": 2.7240332108843811e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4936398749096289e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 294, + "real_time": 2.3905540932109263e+06, + "cpu_time": 2.7283225136054233e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4829129519486301e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 294, + "real_time": 2.3923179191746274e+06, + "cpu_time": 2.7293021598638748e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4788704690729856e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 294, + "real_time": 2.3799439192097001e+06, + "cpu_time": 2.7175765714286347e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5073566625689499e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 294, + "real_time": 2.3911097570050028e+06, + "cpu_time": 2.7291552857142310e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4816387920299791e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3879623935450199e+06, + "cpu_time": 2.7256779482993097e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4888837501060344e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time_median", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3905540932109263e+06, + "cpu_time": 2.7283225136054233e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4829129519486301e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1039100004404518e+03, + "cpu_time": 5.0139300698294128e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1750264499416598e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1373494047632405e-03, + "cpu_time": 1.8395166871999170e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1407384514546524e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 281, + "real_time": 2.4985167797774184e+06, + "cpu_time": 2.8702029572953796e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0491984769594112e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 281, + "real_time": 2.5008234578214614e+06, + "cpu_time": 2.8687805017794291e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0482307304825152e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 281, + "real_time": 2.5009373318920992e+06, + "cpu_time": 2.8679313701066920e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0481830018574411e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 281, + "real_time": 2.4964965654543708e+06, + "cpu_time": 2.8706479857651484e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0500475090871550e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 281, + "real_time": 2.4959243331899955e+06, + "cpu_time": 2.8677088434163732e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0502882499845600e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4985396936270692e+06, + "cpu_time": 2.8690543316726047e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0491895936742166e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time_median", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4985167797774184e+06, + "cpu_time": 2.8687805017794291e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0491984769594112e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3441019673415281e+03, + "cpu_time": 1.3233894299484175e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.8435120295045097e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.3818880417250946e-04, + "cpu_time": 4.6126328642124610e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.3820145461345609e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 269, + "real_time": 2.5941691591601721e+06, + "cpu_time": 3.0297380855018822e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0210247205688441e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 269, + "real_time": 2.5982239030883214e+06, + "cpu_time": 3.0299811672862140e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0178707438447341e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 269, + "real_time": 2.5924727897121673e+06, + "cpu_time": 3.0274019293679786e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0223471663060725e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 269, + "real_time": 2.5930211967598218e+06, + "cpu_time": 3.0257466133829239e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0219194530886903e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 269, + "real_time": 2.5992855926470133e+06, + "cpu_time": 3.0310086728624245e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0170465357217062e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5954345282734991e+06, + "cpu_time": 3.0287752936802851e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0200417239060095e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time_median", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5941691591601721e+06, + "cpu_time": 3.0297380855018822e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0210247205688441e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1148247126153228e+03, + "cpu_time": 2.1463990599411782e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4234026013025569e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2001168508331919e-03, + "cpu_time": 7.0866896742711972e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1996794782122605e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 253, + "real_time": 2.7581646195484009e+06, + "cpu_time": 3.3309661897233543e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8017165203565156e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 253, + "real_time": 2.7534290991738262e+06, + "cpu_time": 3.3236446561265164e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8082549513064563e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 253, + "real_time": 2.7491027377030477e+06, + "cpu_time": 3.3179343162055332e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8142481385621643e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 253, + "real_time": 2.7522895045269267e+06, + "cpu_time": 3.3246407351778592e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8098317719677275e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 253, + "real_time": 2.7547344646376111e+06, + "cpu_time": 3.3229297944664364e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8064503619514614e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7535440851179631e+06, + "cpu_time": 3.3240231383399395e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8081003488288653e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time_median", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7534290991738262e+06, + "cpu_time": 3.3236446561265169e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8082549513064563e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3196047545330730e+03, + "cpu_time": 4.6636839772774902e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5905546854383510e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2055753065565535e-03, + "cpu_time": 1.4030239210689112e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2054710393464604e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 227, + "real_time": 3.0766427732331245e+06, + "cpu_time": 3.9681082466960168e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8163649619815445e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 227, + "real_time": 3.0772616268270998e+06, + "cpu_time": 3.9671290132158580e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8149941549244535e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 227, + "real_time": 3.0781751055483776e+06, + "cpu_time": 3.9665726475770860e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8129717384170449e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 227, + "real_time": 3.0785141182480799e+06, + "cpu_time": 3.9703792599119716e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8122214790863025e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 227, + "real_time": 3.0720971210855613e+06, + "cpu_time": 3.9594586123348013e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8264508488551521e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0765381489884490e+06, + "cpu_time": 3.9663295559471473e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8166006366529000e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time_median", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0772616268270998e+06, + "cpu_time": 3.9671290132158576e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8149941549244535e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5904235432632504e+03, + "cpu_time": 4.1070650884311708e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7447569549981831e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.4199298621243134e-04, + "cpu_time": 1.0354825615216477e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4275979497883341e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 194, + "real_time": 3.6071348088054159e+06, + "cpu_time": 5.0209790927835423e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1627799409551423e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 194, + "real_time": 3.6079892252581483e+06, + "cpu_time": 5.0217967989690630e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1625045802901204e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 194, + "real_time": 3.6080173269567108e+06, + "cpu_time": 5.0233247164948247e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1624955259119585e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 194, + "real_time": 3.6206409051899137e+06, + "cpu_time": 5.0325459896907359e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1584424166417012e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 194, + "real_time": 3.6015314345257645e+06, + "cpu_time": 5.0145468144329553e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1645890300419631e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6090627401471906e+06, + "cpu_time": 5.0226386824742239e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1621622987681770e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time_median", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6079892252581478e+06, + "cpu_time": 5.0217967989690630e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1625045802901204e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0131432386233664e+03, + "cpu_time": 6.4723781006100298e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2547852754530753e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9432034695903752e-03, + "cpu_time": 1.2886409932680331e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9401638461710672e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 138, + "real_time": 4.8973164656132031e+06, + "cpu_time": 7.5690452971014688e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7128989026747828e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 138, + "real_time": 4.8737393833858809e+06, + "cpu_time": 7.5302403550723195e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7211851804378328e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 138, + "real_time": 4.9300065965972086e+06, + "cpu_time": 7.5912432318842215e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7015409281176193e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 138, + "real_time": 4.9261167220285404e+06, + "cpu_time": 7.5886557536232499e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7028845383398933e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 138, + "real_time": 4.9254141572484933e+06, + "cpu_time": 7.5862600362319285e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7031274390712690e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9105186649746653e+06, + "cpu_time": 7.5730889347826391e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7083273977282796e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time_median", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9254141572484933e+06, + "cpu_time": 7.5862600362319294e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7031274390712690e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4351427961404650e+04, + "cpu_time": 2.5483985433426311e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4991424833897222e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.9590337849841539e-03, + "cpu_time": 3.3650714593328281e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9751250812296371e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 88, + "real_time": 7.9415342876348980e+06, + "cpu_time": 1.3688444999999983e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1125912691861579e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 88, + "real_time": 7.9442997653545300e+06, + "cpu_time": 1.3687478920454573e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1118558583559797e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 88, + "real_time": 7.9476107335226098e+06, + "cpu_time": 1.3689998147727262e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1109760609228849e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 88, + "real_time": 7.9379287463697521e+06, + "cpu_time": 1.3695520477272667e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1135508438108258e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 88, + "real_time": 7.9541562008671463e+06, + "cpu_time": 1.3702766090909015e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1092389407905996e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.9451059467497887e+06, + "cpu_time": 1.3692841727272701e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1116425946132898e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time_median", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.9442997653545309e+06, + "cpu_time": 1.3689998147727260e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1118558583559797e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1866780045174401e+03, + "cpu_time": 6.3609265752197298e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6438292734709405e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.7867784847454526e-04, + "cpu_time": 4.6454393484665515e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.7845999018218278e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52, + "real_time": 1.3573708514181467e+07, + "cpu_time": 2.5007326826923307e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4720165432271643e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52, + "real_time": 1.3562874635681510e+07, + "cpu_time": 2.4994525519230790e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4739911634753489e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52, + "real_time": 1.3582966970996214e+07, + "cpu_time": 2.5031430865384918e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4703315609652123e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52, + "real_time": 1.3542523553881507e+07, + "cpu_time": 2.4968074192308072e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4777089636578665e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52, + "real_time": 1.3551415768093787e+07, + "cpu_time": 2.4973335019230746e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4760831321405129e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3562697888566900e+07, + "cpu_time": 2.4994938484615572e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4740262726932211e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time_median", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3562874635681510e+07, + "cpu_time": 2.4994525519230794e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4739911634753489e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6327820128425297e+04, + "cpu_time": 2.5854087377625277e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9784233859946774e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2038770060777764e-03, + "cpu_time": 1.0343729148818875e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2038770238087934e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 28, + "real_time": 2.4621350025492053e+07, + "cpu_time": 4.7668973642857842e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7256370560719829e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 28, + "real_time": 2.4531674238720112e+07, + "cpu_time": 4.7665280607142381e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7356006502840819e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 28, + "real_time": 2.4622675829700060e+07, + "cpu_time": 4.7666441000000000e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7254902945622497e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 28, + "real_time": 2.4540257307567768e+07, + "cpu_time": 4.7663883714285754e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7346438612648473e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 28, + "real_time": 2.4669779198510308e+07, + "cpu_time": 4.7676848035714723e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7202863657593007e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4597147319998063e+07, + "cpu_time": 4.7668285400000140e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7283316455884924e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time_median", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4621350025492057e+07, + "cpu_time": 4.7666441000000000e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7256370560719829e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.9236946712670629e+04, + "cpu_time": 5.1380403583964617e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5715886942260778e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4082852349511926e-03, + "cpu_time": 1.0778739607018560e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4086473156047007e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 16, + "real_time": 4.4960850849747658e+07, + "cpu_time": 8.7076805937499553e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9852132569406943e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 16, + "real_time": 4.4936058809980750e+07, + "cpu_time": 8.6973259249999657e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9868602533115096e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 16, + "real_time": 4.4920268934220076e+07, + "cpu_time": 8.6941317437499776e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9879101613693476e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 16, + "real_time": 4.4870279263705015e+07, + "cpu_time": 8.6925117750000775e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9912389715962157e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 16, + "real_time": 4.4915589503943920e+07, + "cpu_time": 8.6930255500000447e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9882214501095371e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4920609472319491e+07, + "cpu_time": 8.6969351175000042e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9878888186654611e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time_median", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4920268934220076e+07, + "cpu_time": 8.6941317437499776e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9879101613693476e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3228363016630348e+04, + "cpu_time": 6.2914142223657618e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2108370063846861e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.3971309398876222e-04, + "cpu_time": 7.2340590534085468e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3993282232374209e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.5369773209095001e+07, + "cpu_time": 1.6439538887500051e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1443852538125496e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.5412476211786270e+07, + "cpu_time": 1.6441444587499630e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1428131802945895e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.5332995280623436e+07, + "cpu_time": 1.6439301562500221e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1457404620244665e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.3763759583234787e+07, + "cpu_time": 1.6052745387500522e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.2046729675888023e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.6012801155447960e+07, + "cpu_time": 1.6489506962499690e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1208779669303627e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.5178361088037506e+07, + "cpu_time": 1.6372507477500021e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1516979661301546e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time_median", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.5369773209095001e+07, + "cpu_time": 1.6439538887500051e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1443852538125496e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.3856244076576899e+05, + "cpu_time": 1.8003031588035836e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1321439008127324e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.8447825252127013e-03, + "cpu_time": 1.0995891504570431e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9379570456066511e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6405433043837547e+08, + "cpu_time": 3.0734573049998915e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2725189915158467e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6685758903622627e+08, + "cpu_time": 3.1158725099999171e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2175396702120667e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6408236697316170e+08, + "cpu_time": 3.0733097300000620e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2719598205687380e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.6688183322548866e+08, + "cpu_time": 3.1161905550000310e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2170722338277931e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6372112184762955e+08, + "cpu_time": 3.0677975950000304e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2791792893995070e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6511944830417636e+08, + "cpu_time": 3.0893255389999872e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2516540011047902e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time_median", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6408236697316170e+08, + "cpu_time": 3.0734573049998915e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2719598205687380e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6040935485117952e+06, + "cpu_time": 2.4485863855554787e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1484130228306450e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.7147462941906109e-03, + "cpu_time": 7.9259578009641708e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6824970361573896e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.1946991384029388e+08, + "cpu_time": 5.6197194700001776e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3610107790518703e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.1875278055667877e+08, + "cpu_time": 5.6147966449998426e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3685724156657934e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.1801357865333557e+08, + "cpu_time": 5.6053470300000191e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3764024434015713e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.1874062120914459e+08, + "cpu_time": 5.6025327500000799e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3687009202867007e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.1900304555892944e+08, + "cpu_time": 5.6124025049999201e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3659296954945455e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1879598796367645e+08, + "cpu_time": 5.6109596800000083e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3681232507800965e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time_median", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1875278055667877e+08, + "cpu_time": 5.6124025049999213e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3685724156657934e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2776548326815199e+05, + "cpu_time": 7.0009455513923254e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5785172870503860e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6554960011864561e-03, + "cpu_time": 1.2477269398935185e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6562687501885023e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2740403996091094e+06, + "cpu_time": 2.5802008636364462e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8011993105769237e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2664793227291242e+06, + "cpu_time": 2.5728709545453717e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8072081924259095e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2722938931063302e+06, + "cpu_time": 2.5853760551948557e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8025837293434697e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2670486760507156e+06, + "cpu_time": 2.5734621590910084e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8067543248058469e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2666424016964128e+06, + "cpu_time": 2.5796687629869590e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8070781685432382e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2693009386383384e+06, + "cpu_time": 2.5783157590909279e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8049647451390778e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time_median", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2670486760507156e+06, + "cpu_time": 2.5796687629869585e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8067543248058469e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5889513966643208e+03, + "cpu_time": 5.2066147652994805e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8526278785706754e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5815228978919914e-03, + "cpu_time": 2.0193860069083421e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5804341255157713e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 306, + "real_time": 2.2680473329767194e+06, + "cpu_time": 2.5767117581698825e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6119175648985836e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 306, + "real_time": 2.2680457746853530e+06, + "cpu_time": 2.5838065032681189e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6119200465151458e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 306, + "real_time": 2.2759339079759968e+06, + "cpu_time": 2.5844830980393118e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.5994015341531602e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 306, + "real_time": 2.2678773764983499e+06, + "cpu_time": 2.5826587908495953e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6121882447844772e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 306, + "real_time": 2.2676912398120156e+06, + "cpu_time": 2.5826906862744703e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6124847405059827e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2695191263896869e+06, + "cpu_time": 2.5820701673202761e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6095824261714700e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time_median", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2680457746853526e+06, + "cpu_time": 2.5826906862744703e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6119200465151458e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5889718562761968e+03, + "cpu_time": 3.0935888233694636e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6960868842967129e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5813798678953953e-03, + "cpu_time": 1.1981040881549909e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5780459376677288e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2755634952820353e+06, + "cpu_time": 2.5958982532468727e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1999748782968381e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2822175539316391e+06, + "cpu_time": 2.5937507175324773e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1789825522000873e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2805583324622023e+06, + "cpu_time": 2.5923543311688933e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1842056249054726e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2814482111822474e+06, + "cpu_time": 2.5934014253245061e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1814034259887077e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2750845286305752e+06, + "cpu_time": 2.5876292077921792e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2014906671893634e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2789744242977398e+06, + "cpu_time": 2.5926067870129859e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1892114297160944e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time_median", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2805583324622018e+06, + "cpu_time": 2.5934014253245061e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1842056249054726e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3879197708770944e+03, + "cpu_time": 3.0665047650270394e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0692105002056423e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4865984167071436e-03, + "cpu_time": 1.1827882193273299e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4872430873101557e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 306, + "real_time": 2.2883130461160168e+06, + "cpu_time": 2.6101910784313167e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4319719085471086e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 306, + "real_time": 2.2888296425306038e+06, + "cpu_time": 2.6055486111110123e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4316487077548787e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 306, + "real_time": 2.2940490116346991e+06, + "cpu_time": 2.6167885686275032e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4283914525718914e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 306, + "real_time": 2.2895977462587110e+06, + "cpu_time": 2.6061343823528946e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4311684248267690e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 306, + "real_time": 2.3021671983407312e+06, + "cpu_time": 2.6183921013071560e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4233544819688717e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2925913289761525e+06, + "cpu_time": 2.6114109483659766e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4293069951339040e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time_median", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2895977462587110e+06, + "cpu_time": 2.6101910784313167e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4311684248267690e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.8145859580452106e+03, + "cpu_time": 5.9445884150628754e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6166605163098146e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5362505233943915e-03, + "cpu_time": 2.2763894816258423e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5303594879356127e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 301, + "real_time": 2.3247020031353366e+06, + "cpu_time": 2.6494600631228671e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8191140159732856e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 301, + "real_time": 2.3260978653878467e+06, + "cpu_time": 2.6644881096346602e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8174223008915711e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 301, + "real_time": 2.3251096942335167e+06, + "cpu_time": 2.6498823887042692e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8186197048051208e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 301, + "real_time": 2.3363363415473127e+06, + "cpu_time": 2.6618034750831355e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8050755721497152e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 301, + "real_time": 2.3238382037679520e+06, + "cpu_time": 2.6489859900331274e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8201619154783521e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3272168216143930e+06, + "cpu_time": 2.6549240053156121e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8160787018596090e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time_median", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3251096942335172e+06, + "cpu_time": 2.6498823887042692e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8186197048051208e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1622926877957962e+03, + "cpu_time": 7.5718522097648483e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2292973714687520e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2182259254273987e-03, + "cpu_time": 2.8520033698157476e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2120466190647336e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 294, + "real_time": 2.3802082651440487e+06, + "cpu_time": 2.7382729319727481e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5067450155277744e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 294, + "real_time": 2.3846621490495442e+06, + "cpu_time": 2.7306570918367039e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4964599514543980e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 294, + "real_time": 2.3795628158350280e+06, + "cpu_time": 2.7312209931972972e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5082387036714837e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 294, + "real_time": 2.3794136846400038e+06, + "cpu_time": 2.7292342142857253e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5085839358711891e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 294, + "real_time": 2.3720187949156072e+06, + "cpu_time": 2.7162866020408855e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5257572275966451e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3791731419168469e+06, + "cpu_time": 2.7291343666666723e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5091569668242984e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time_median", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3795628158350280e+06, + "cpu_time": 2.7306570918367039e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5082387036714837e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5440914100979153e+03, + "cpu_time": 7.9884463992943956e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0532947566445966e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9099456571861106e-03, + "cpu_time": 2.9270989720639427e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9118982504718112e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 278, + "real_time": 2.4999116284424346e+06, + "cpu_time": 2.8717994424459329e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0486130670280068e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 278, + "real_time": 2.4887081244022404e+06, + "cpu_time": 2.8692787122301147e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0533336449928777e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 278, + "real_time": 2.4880450545548857e+06, + "cpu_time": 2.8634275971221407e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0536143608818124e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 278, + "real_time": 2.4948900284098205e+06, + "cpu_time": 2.8663096546762986e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0507236672354810e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 278, + "real_time": 2.4954149358832152e+06, + "cpu_time": 2.8705620899280091e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0505026487998399e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4933939543385194e+06, + "cpu_time": 2.8682754992804993e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0513574777876037e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time_median", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4948900284098210e+06, + "cpu_time": 2.8692787122301143e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0507236672354810e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9842354187181409e+03, + "cpu_time": 3.3918950743710111e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1013684205179254e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9989762989701420e-03, + "cpu_time": 1.1825555373679623e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9987192414704508e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 270, + "real_time": 2.5931684360666960e+06, + "cpu_time": 3.0372282481481405e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0218046491235149e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 270, + "real_time": 2.5942422035667631e+06, + "cpu_time": 3.0367456037037368e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0209678158776718e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 270, + "real_time": 2.5983737574683297e+06, + "cpu_time": 3.0422832407408706e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0177543684509379e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 270, + "real_time": 2.5916649822007725e+06, + "cpu_time": 3.0356337037037266e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0229775206315002e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 270, + "real_time": 2.5979542113081724e+06, + "cpu_time": 3.0377084037038712e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0180802175724271e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5950807181221466e+06, + "cpu_time": 3.0379198400000688e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0203169143312106e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time_median", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5942422035667626e+06, + "cpu_time": 3.0372282481481405e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0209678158776718e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9634545902166024e+03, + "cpu_time": 2.5577099456921746e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3068001265822750e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1419508339459354e-03, + "cpu_time": 8.4192805617021055e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1418011254664467e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 254, + "real_time": 2.7590429641335732e+06, + "cpu_time": 3.3365889803149472e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8005062394136584e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 254, + "real_time": 2.7523954054267388e+06, + "cpu_time": 3.3357813228346505e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8096851852483964e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 254, + "real_time": 2.7559425967826149e+06, + "cpu_time": 3.3325873818898979e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8047817150623703e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 254, + "real_time": 2.7557339805259011e+06, + "cpu_time": 3.3370958385827234e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8050697469713348e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 254, + "real_time": 2.7534939920691056e+06, + "cpu_time": 3.3305816102361991e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8081652003607625e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7553217877875869e+06, + "cpu_time": 3.3345270267716842e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8056416174113053e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time_median", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7557339805259006e+06, + "cpu_time": 3.3357813228346501e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8050697469713348e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5645180334510442e+03, + "cpu_time": 2.8174562572764989e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5410975264739600e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.3075082729638259e-04, + "cpu_time": 8.4493429942423170e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.3048633646241892e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 226, + "real_time": 3.0876570000041244e+06, + "cpu_time": 3.9918503362831874e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7920497645858943e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 226, + "real_time": 3.0931550968207614e+06, + "cpu_time": 4.0049350486724889e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7799768661956728e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 226, + "real_time": 3.0936124565623766e+06, + "cpu_time": 4.0098563982301992e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7789745142491317e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 226, + "real_time": 3.0891883529030378e+06, + "cpu_time": 3.9950284955752213e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7886828526632881e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 226, + "real_time": 3.0823281257355637e+06, + "cpu_time": 3.9915204247788815e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8037921806249547e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0891882064051731e+06, + "cpu_time": 3.9986381407079957e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7886952356637883e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time_median", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0891883529030378e+06, + "cpu_time": 3.9950284955752208e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7886828526632881e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6020514142857373e+03, + "cpu_time": 8.2955719921366290e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0120192140073231e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4897284033209010e-03, + "cpu_time": 2.0745993261265250e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4907418566837010e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 193, + "real_time": 3.6158287340360139e+06, + "cpu_time": 5.0643369222795600e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1599841443038392e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 193, + "real_time": 3.6166701622674037e+06, + "cpu_time": 5.0666559378239252e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1597142708115964e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 193, + "real_time": 3.6216634048216548e+06, + "cpu_time": 5.0716556217615977e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1581153550647385e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 193, + "real_time": 3.6184036841258947e+06, + "cpu_time": 5.0694190569948694e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1591586694432704e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 193, + "real_time": 3.6164578955157278e+06, + "cpu_time": 5.0680559326425744e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1597823398416388e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6178047761533386e+06, + "cpu_time": 5.0680246943005044e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1593509558930166e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time_median", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6166701622674041e+06, + "cpu_time": 5.0680559326425735e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1597142708115964e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3587011540937087e+03, + "cpu_time": 2.7659543634119300e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5543048834422883e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.5197026927517570e-04, + "cpu_time": 5.4576576284691739e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5159776209641469e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 142, + "real_time": 5.0026962671323977e+06, + "cpu_time": 7.7243923239435190e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6768173704873841e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 142, + "real_time": 4.9915525727761043e+06, + "cpu_time": 7.7089918943661656e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6805608831511490e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 142, + "real_time": 4.9842680365004586e+06, + "cpu_time": 7.7098018661971176e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6830170325048945e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 142, + "real_time": 4.9867818236980643e+06, + "cpu_time": 7.7096942042253837e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6821686403314979e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 142, + "real_time": 4.9806601256394470e+06, + "cpu_time": 7.7086792605634509e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6842361832354541e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9891917651492944e+06, + "cpu_time": 7.7123119098591283e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6813600219420762e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time_median", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9867818236980643e+06, + "cpu_time": 7.7096942042253856e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6821686403314979e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.5262063781610177e+03, + "cpu_time": 6.7695596268210647e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8699371305727237e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7089353906415508e-03, + "cpu_time": 8.7776009398259883e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7069141011559002e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 87, + "real_time": 7.8310137668132093e+06, + "cpu_time": 1.3328289494253093e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1424066537974432e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 87, + "real_time": 7.9148305237464523e+06, + "cpu_time": 1.3469970114942562e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1197189187644885e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 87, + "real_time": 7.8411712257683966e+06, + "cpu_time": 1.3342888264367569e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1396313786472526e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 87, + "real_time": 7.9504868668643227e+06, + "cpu_time": 1.3521041390804147e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1102124034596319e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 87, + "real_time": 7.9320114287922438e+06, + "cpu_time": 1.3509693310344698e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1151275626130254e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8939027623969261e+06, + "cpu_time": 1.3434376514942413e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1254193834563684e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time_median", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.9148305237464514e+06, + "cpu_time": 1.3469970114942562e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1197189187644885e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4377484422839341e+04, + "cpu_time": 9.2296913126719825e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4664730440247444e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.8885424687354533e-03, + "cpu_time": 6.8702044359157562e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8996879177791164e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.3716290467510037e+07, + "cpu_time": 2.5098387529411349e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4463197305044565e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.3791783000616467e+07, + "cpu_time": 2.5200691117647115e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4329292302888017e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.3774894926623970e+07, + "cpu_time": 2.5180097196078029e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4359120108528996e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3759719938331960e+07, + "cpu_time": 2.5136025549019735e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4385984707816429e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.3808495145948494e+07, + "cpu_time": 2.5239101372549426e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4299847047304859e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3770236695806187e+07, + "cpu_time": 2.5170860552941132e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4367488294316573e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time_median", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3774894926623970e+07, + "cpu_time": 2.5180097196078025e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4359120108528996e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5250706047941683e+04, + "cpu_time": 5.4995009468825046e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2463250473986706e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5599201253147311e-03, + "cpu_time": 2.1848680681042132e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5633848560647722e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 28, + "real_time": 2.4365937137710195e+07, + "cpu_time": 4.6527072500000238e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.7542082055255027e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 28, + "real_time": 2.5414041376539640e+07, + "cpu_time": 4.8346264250000104e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6406214976084023e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 28, + "real_time": 2.5414927396923304e+07, + "cpu_time": 4.8319043928572118e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6405294397230544e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 28, + "real_time": 2.5380578424249377e+07, + "cpu_time": 4.8316446499999948e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6441030176003456e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 28, + "real_time": 2.5391927720712762e+07, + "cpu_time": 4.8314613500001058e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6429211967730122e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5193482411227059e+07, + "cpu_time": 4.7964688135714695e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6644766714460640e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time_median", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5391927720712766e+07, + "cpu_time": 4.8316446499999955e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6429211967730122e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6284467630482570e+05, + "cpu_time": 8.0375505945367506e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0184688137070656e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8371603764415143e-02, + "cpu_time": 1.6757224756251379e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8834726036401902e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15, + "real_time": 4.5995108286539711e+07, + "cpu_time": 8.8087951533335239e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9180870096848602e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15, + "real_time": 4.5896399269501366e+07, + "cpu_time": 8.7993801066666797e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9243629159638472e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15, + "real_time": 4.6128623684247337e+07, + "cpu_time": 8.8201294733335078e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9096408537728515e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15, + "real_time": 4.6641982346773148e+07, + "cpu_time": 8.9244664399999842e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8776162857341685e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15, + "real_time": 4.6054171770811081e+07, + "cpu_time": 8.8226635599998340e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9143446258883014e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6143257071574524e+07, + "cpu_time": 8.8350869466667056e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9088103382088060e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time_median", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6054171770811081e+07, + "cpu_time": 8.8201294733335093e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9143446258883014e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9144642042298720e+05, + "cpu_time": 5.0825675580236793e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8250522694612339e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.3161215509974465e-03, + "cpu_time": 5.7527080250650226e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2742223014274241e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.7171018123626709e+07, + "cpu_time": 1.6666201262500378e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0794117331439500e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.7379561737179756e+07, + "cpu_time": 1.6679597499999943e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0720622839400382e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.7991042993962765e+07, + "cpu_time": 1.6819663249999905e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0507134233926272e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.7203377857804298e+07, + "cpu_time": 1.6671786112500086e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0782690142773671e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.7428119033575058e+07, + "cpu_time": 1.6681861612499917e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0703560704184041e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7434623949229732e+07, + "cpu_time": 1.6703821947500044e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0701625050344772e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time_median", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7379561737179756e+07, + "cpu_time": 1.6679597499999943e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0720622839400382e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3001481048313581e+05, + "cpu_time": 6.5057471128992795e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1545314167304413e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7744179087996540e-03, + "cpu_time": 3.8947656011580943e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7604895989617206e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6999528184533119e+08, + "cpu_time": 3.1636755674999505e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1581518391108499e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6916919872164726e+08, + "cpu_time": 3.1405298949999863e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1735736532237935e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6966279968619347e+08, + "cpu_time": 3.1576139750001174e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1643407570368447e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.6935619711875916e+08, + "cpu_time": 3.1438360450000858e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1700694815644994e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6979407519102097e+08, + "cpu_time": 3.1618111724999666e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1618942615990100e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6959551051259041e+08, + "cpu_time": 3.1534933310000217e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1656059985070000e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time_median", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6966279968619347e+08, + "cpu_time": 3.1576139750001174e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1643407570368447e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3271065749994590e+05, + "cpu_time": 1.0620239876626271e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2118540273898905e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9617892979262924e-03, + "cpu_time": 3.3677698862481596e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9622953805115348e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.2319845259189606e+08, + "cpu_time": 5.6536951800001132e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3222368962137885e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.2301260530948639e+08, + "cpu_time": 5.6463897800000501e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3241483655760779e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.2289998233318329e+08, + "cpu_time": 5.6270940050001168e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3253077818135743e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.2324321568012238e+08, + "cpu_time": 5.6498326799999177e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3217768290690503e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.2389615476131439e+08, + "cpu_time": 5.6622772250000250e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3150804917436023e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2325008213520050e+08, + "cpu_time": 5.6478577740000451e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3217100728832188e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time_median", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2319845259189606e+08, + "cpu_time": 5.6498326799999177e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3222368962137885e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8700378236959933e+05, + "cpu_time": 1.3030591305971283e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9724717148294463e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1972271741225223e-03, + "cpu_time": 2.3071741228962439e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1959116321616147e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2617678450342137e+06, + "cpu_time": 2.5706053527507298e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8109727790997224e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2614665343019278e+06, + "cpu_time": 2.5816373009708682e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8112140674521888e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2619253010895937e+06, + "cpu_time": 2.5709759902913468e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8108467145342568e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2615809618980656e+06, + "cpu_time": 2.5772811747574201e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8111224267480439e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2621702328987396e+06, + "cpu_time": 2.5766974919094499e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8106506488467914e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2617821750445077e+06, + "cpu_time": 2.5754394621359631e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8109613273362007e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time_median", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2617678450342133e+06, + "cpu_time": 2.5766974919094495e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8109727790997224e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7901394225090155e+02, + "cpu_time": 4.6550763949436569e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2339368150974315e+02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2336021803046135e-04, + "cpu_time": 1.8074881834275109e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2335640642218454e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2706912143282518e+06, + "cpu_time": 2.5824689288026951e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6077120254430869e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2635139543020609e+06, + "cpu_time": 2.5824472038834020e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6191515340253101e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2631157671845849e+06, + "cpu_time": 2.5810856537216669e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6197883107814710e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2636299597213953e+06, + "cpu_time": 2.5817393333332157e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6189660614883630e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2632964957553688e+06, + "cpu_time": 2.5814587508091372e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6194992637347514e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2648494782583322e+06, + "cpu_time": 2.5818399741100236e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6170234390945965e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time_median", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2635139543020604e+06, + "cpu_time": 2.5817393333332157e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6191515340253101e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2716295157215650e+03, + "cpu_time": 6.1007602247473039e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2148637824989764e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4445240388502312e-03, + "cpu_time": 2.3629505646841163e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4417555955358551e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2698396532614897e+06, + "cpu_time": 2.5899726623376552e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2181310148750553e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2776438552257302e+06, + "cpu_time": 2.5926029805195378e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1933985475425581e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2755108679971332e+06, + "cpu_time": 2.5893735811688905e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2001413970045876e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2698900626077279e+06, + "cpu_time": 2.5899391948051411e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2179707158052828e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2903421346412655e+06, + "cpu_time": 2.6046404253247674e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1535163904960481e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2766453147466695e+06, + "cpu_time": 2.5933057688311981e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.1966316131447069e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time_median", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2755108679971332e+06, + "cpu_time": 2.5899726623376552e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2001413970045876e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.3940066036735698e+03, + "cpu_time": 6.4589834459351796e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6454088989560350e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6870067328022074e-03, + "cpu_time": 2.4906370562103983e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6758987275716240e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 306, + "real_time": 2.2890830741209141e+06, + "cpu_time": 2.6133458398691737e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4314902054214012e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 306, + "real_time": 2.2965932190357372e+06, + "cpu_time": 2.6154015947712669e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4268090547510276e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 306, + "real_time": 2.2942919970423081e+06, + "cpu_time": 2.6119502712417506e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4282401735368883e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 306, + "real_time": 2.2877266849665083e+06, + "cpu_time": 2.6115458333333437e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4323389334630994e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 306, + "real_time": 2.2880054235738381e+06, + "cpu_time": 2.6061423006537738e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4321644373035079e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2911400797478617e+06, + "cpu_time": 2.6116771679738620e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4302085608951850e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time_median", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2890830741209141e+06, + "cpu_time": 2.6119502712417510e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4314902054214012e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0428924092941952e+03, + "cpu_time": 3.4417441540627842e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5217921856755715e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7645767035506239e-03, + "cpu_time": 1.3178290932232208e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7632338769508909e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 302, + "real_time": 2.3251949818743183e+06, + "cpu_time": 2.6515963443709752e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8185163184539489e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 302, + "real_time": 2.3194566470823740e+06, + "cpu_time": 2.6543156158940797e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8254893266678303e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 302, + "real_time": 2.3195504076741883e+06, + "cpu_time": 2.6534715662252577e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8253751150729638e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 302, + "real_time": 2.3186789310555784e+06, + "cpu_time": 2.6465213807946146e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8264370337019775e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 302, + "real_time": 2.3242635706390273e+06, + "cpu_time": 2.6516844768213513e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8196457935268372e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3214289076650972e+06, + "cpu_time": 2.6515178768212562e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8230927174847115e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time_median", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3195504076741887e+06, + "cpu_time": 2.6516844768213518e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8253751150729638e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0495726955644100e+03, + "cpu_time": 3.0266785202288488e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7068260736949880e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3136618939718825e-03, + "cpu_time": 1.1414890115156796e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3130373121424278e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 293, + "real_time": 2.3914142551517221e+06, + "cpu_time": 2.7461987576792547e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4809408164075777e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 293, + "real_time": 2.3992116993719097e+06, + "cpu_time": 2.7508107372013074e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4631277445968352e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 293, + "real_time": 2.3897673192500761e+06, + "cpu_time": 2.7448081877132640e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4847180704241619e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 293, + "real_time": 2.3925845247435896e+06, + "cpu_time": 2.7387553754267558e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4782599588220119e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 293, + "real_time": 2.3878303677504789e+06, + "cpu_time": 2.7429817918089614e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4891671439575493e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3921616332535557e+06, + "cpu_time": 2.7447109699659087e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4792427468416274e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time_median", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3914142551517221e+06, + "cpu_time": 2.7448081877132640e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4809408164075777e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3282052838400177e+03, + "cpu_time": 4.4127453842625137e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.8996898480457457e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8093281088006031e-03, + "cpu_time": 1.6077267998521982e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8067624132462783e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 280, + "real_time": 2.5130426003930294e+06, + "cpu_time": 2.8887264714287296e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0431339284061551e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 280, + "real_time": 2.5093534296112400e+06, + "cpu_time": 2.8854687250001477e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0446675103897679e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 280, + "real_time": 2.5041013648400879e+06, + "cpu_time": 2.8830379821428666e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0468585804103042e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 280, + "real_time": 2.5032288112145448e+06, + "cpu_time": 2.8867165964285848e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0472234852267061e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 280, + "real_time": 2.5090242769303066e+06, + "cpu_time": 2.8841511392857102e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0448045577331877e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5077500965978419e+06, + "cpu_time": 2.8856201828572075e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0453376124332243e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time_median", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5090242769303066e+06, + "cpu_time": 2.8854687250001477e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0448045577331877e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0607903680286208e+03, + "cpu_time": 2.2192307468903091e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6925769653105881e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6192962662179629e-03, + "cpu_time": 7.6906543698101310e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6191677647289374e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 269, + "real_time": 2.5998854067965955e+06, + "cpu_time": 3.0546604237919413e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0165811871146756e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 269, + "real_time": 2.5929672994952109e+06, + "cpu_time": 3.0412316394053753e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0219614805866098e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 269, + "real_time": 2.5894383538071220e+06, + "cpu_time": 3.0425800074348291e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0247170558401808e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 269, + "real_time": 2.5943757742851088e+06, + "cpu_time": 3.0432616171002984e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0208637669092861e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 269, + "real_time": 2.5928786697842483e+06, + "cpu_time": 3.0486680594795793e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0220305952211240e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5939091008336572e+06, + "cpu_time": 3.0460803494424052e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0212308171343753e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time_median", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5929672994952113e+06, + "cpu_time": 3.0432616171002984e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0219614805866098e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8036289890873354e+03, + "cpu_time": 5.5682482498236041e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9614675233741803e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4663694220683700e-03, + "cpu_time": 1.8280043895896869e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4651802744492274e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 255, + "real_time": 2.7545326654160139e+06, + "cpu_time": 3.3371761843136037e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8067292254878187e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 255, + "real_time": 2.7469868926952281e+06, + "cpu_time": 3.3317379137253170e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8171860331345856e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 255, + "real_time": 2.7502471617623870e+06, + "cpu_time": 3.3316905254902812e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8126609658168375e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 255, + "real_time": 2.7496761870661788e+06, + "cpu_time": 3.3312826000002022e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8134526710172331e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 255, + "real_time": 2.7589530985364141e+06, + "cpu_time": 3.3444051960785538e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8006300308484942e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7520792010952444e+06, + "cpu_time": 3.3352584839215917e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8101317852609944e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time_median", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7502471617623875e+06, + "cpu_time": 3.3317379137253170e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8126609658168375e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6992659690400969e+03, + "cpu_time": 5.6628683742474950e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5009268465350382e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7075329689530488e-03, + "cpu_time": 1.6978799099220351e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7062209952115144e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 227, + "real_time": 3.0880496439232701e+06, + "cpu_time": 3.9960100925111067e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7911861589622450e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 227, + "real_time": 3.0870410072153625e+06, + "cpu_time": 3.9952476123349266e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7934050603743589e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 227, + "real_time": 3.0811194993046806e+06, + "cpu_time": 3.9922086607928895e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8064610946549344e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 227, + "real_time": 3.0849827761199768e+06, + "cpu_time": 3.9942376519824360e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7979374673774207e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 227, + "real_time": 3.0886036218738896e+06, + "cpu_time": 3.9936098590308954e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7899680786090469e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0859593096874366e+06, + "cpu_time": 3.9942627753304513e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7957915719956017e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time_median", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0870410072153625e+06, + "cpu_time": 3.9942376519824364e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7934050603743589e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0376527138733454e+03, + "cpu_time": 1.4721870356938894e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6944365504499327e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.8434632768408612e-04, + "cpu_time": 3.6857540890561298e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.8508561946435537e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 194, + "real_time": 3.6133182494296241e+06, + "cpu_time": 5.0694864742267448e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1607900855846524e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 194, + "real_time": 3.6105479080153187e+06, + "cpu_time": 5.0682160309279589e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1616807495307729e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 194, + "real_time": 3.6116563306986978e+06, + "cpu_time": 5.0710014793815315e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1613242279861619e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 194, + "real_time": 3.6085677061940436e+06, + "cpu_time": 5.0675120567009412e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1623182219362409e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 194, + "real_time": 3.6135252238702527e+06, + "cpu_time": 5.0693598453608742e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1607235981897218e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6115230836415873e+06, + "cpu_time": 5.0691151773196105e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1613673766455102e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time_median", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6116563306986974e+06, + "cpu_time": 5.0693598453608733e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1613242279861619e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0575570366558641e+03, + "cpu_time": 1.3355073238357097e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6179430447836278e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.6972002919642947e-04, + "cpu_time": 2.6345965264531324e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6984061872814723e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 143, + "real_time": 4.8775992786081936e+06, + "cpu_time": 7.5926146713288985e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7198231180634544e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 143, + "real_time": 4.9225596378889531e+06, + "cpu_time": 7.6511889020979600e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7041150574251785e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 143, + "real_time": 4.9156864089044658e+06, + "cpu_time": 7.6432798321677623e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7064977913978705e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 143, + "real_time": 4.9012493394263145e+06, + "cpu_time": 7.6316267062939759e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7115244336828368e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 143, + "real_time": 4.9758792294749431e+06, + "cpu_time": 7.7281422097903294e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6858544215280662e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9185947788605746e+06, + "cpu_time": 7.6493704643357862e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7055629644194813e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time_median", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9156864089044658e+06, + "cpu_time": 7.6432798321677623e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7064977913978705e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6356506691375216e+04, + "cpu_time": 4.9450610176625567e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2552769582602877e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.3916450380564682e-03, + "cpu_time": 6.4646640409407197e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3598980773339172e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 87, + "real_time": 7.8139821463532150e+06, + "cpu_time": 1.3315634505747296e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1470763159895275e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 87, + "real_time": 7.8818503146370249e+06, + "cpu_time": 1.3431383034482950e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1285885078080964e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 87, + "real_time": 7.8420799557419345e+06, + "cpu_time": 1.3356930241379371e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1393834409601755e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 87, + "real_time": 7.0550440136214783e+06, + "cpu_time": 1.2511515413792893e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.3780455469317417e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 87, + "real_time": 7.8308246601587059e+06, + "cpu_time": 1.3351079689655386e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1424583907948184e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6847562181024719e+06, + "cpu_time": 1.3193308577011582e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1871104404968719e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time_median", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8308246601587059e+06, + "cpu_time": 1.3351079689655384e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1424583907948184e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5290566447809618e+05, + "cpu_time": 3.8345429016643402e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0695260004712817e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.5922818429396578e-02, + "cpu_time": 2.9064300886176220e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8901325724927946e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.3754747461016271e+07, + "cpu_time": 2.5071099882353008e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4394800482597027e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.3699989400657954e+07, + "cpu_time": 2.5015458470588829e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4492305080461245e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.3709755870057087e+07, + "cpu_time": 2.5001317666666925e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4474857406676989e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3728400692343712e+07, + "cpu_time": 2.5014852254902478e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4441617601322784e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.3730358387179235e+07, + "cpu_time": 2.5001228686273947e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4438132679283562e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3724650362250853e+07, + "cpu_time": 2.5020791392157041e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4448342650068321e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time_median", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3728400692343712e+07, + "cpu_time": 2.5014852254902482e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4441617601322784e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1109777372287343e+04, + "cpu_time": 2.8968012685471836e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7589762854905059e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5380921783150854e-03, + "cpu_time": 1.1577576516845140e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5375178347641497e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 28, + "real_time": 2.5303560309112072e+07, + "cpu_time": 4.8225034035713829e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6521510483184223e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 28, + "real_time": 2.5361097949956145e+07, + "cpu_time": 4.8305348250002973e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6461340172425795e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 28, + "real_time": 2.5523687579802103e+07, + "cpu_time": 4.8492951107139759e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6292777558171444e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 28, + "real_time": 2.5338337690170322e+07, + "cpu_time": 4.8275923428572498e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6485109173532724e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 28, + "real_time": 2.5271040187882524e+07, + "cpu_time": 4.8171336607144408e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6555639776228414e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5359544743384637e+07, + "cpu_time": 4.8294118685714699e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6463275432708521e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time_median", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5338337690170322e+07, + "cpu_time": 4.8275923428572506e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6485109173532724e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7931526498154140e+04, + "cpu_time": 1.2231070066644539e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0182396052796219e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.8617225777959141e-03, + "cpu_time": 2.5326210311945223e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8477459370773169e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15, + "real_time": 4.7390252600113548e+07, + "cpu_time": 9.0851348933332101e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8321800504536328e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15, + "real_time": 4.5897176116704941e+07, + "cpu_time": 8.8001413400002092e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9243134187323022e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15, + "real_time": 4.5885218183199562e+07, + "cpu_time": 8.8048880933335260e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9250755104645562e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15, + "real_time": 4.5887223631143570e+07, + "cpu_time": 8.7960075799999982e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9249476734283547e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15, + "real_time": 4.5881142963965736e+07, + "cpu_time": 8.8029465933338240e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9253353192489629e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6188202699025467e+07, + "cpu_time": 8.8578237000001535e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9063703944655619e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time_median", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5887223631143570e+07, + "cpu_time": 8.8029465933338240e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9249476734283547e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7199221225961146e+05, + "cpu_time": 1.2711465420123679e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1475367070954300e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4549001108323922e-02, + "cpu_time": 1.4350551388964150e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4270502875316071e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.7700656615197659e+07, + "cpu_time": 1.6708382825000000e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0608146661638880e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.7124891579151154e+07, + "cpu_time": 1.6662926112500998e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0810420665617929e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.7260894477367401e+07, + "cpu_time": 1.6661283925000703e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0762400226097074e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.7061014026403427e+07, + "cpu_time": 1.6663166437500367e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0833026585078626e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.7402153760194778e+07, + "cpu_time": 1.6699241199999905e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0712682062332945e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7309922091662884e+07, + "cpu_time": 1.6679000100000396e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0745335240153093e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time_median", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7260894477367401e+07, + "cpu_time": 1.6663166437500364e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0762400226097074e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5484081847121910e+05, + "cpu_time": 2.2890977786925217e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9587993648434840e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9188070767452163e-03, + "cpu_time": 1.3724430511229911e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9138727208098169e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6787138953804970e+08, + "cpu_time": 3.1163635574998236e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1981084655185566e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6964251548051834e+08, + "cpu_time": 3.1615873999999166e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1647191181956625e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6738265752792358e+08, + "cpu_time": 3.1069012400001836e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2074464578890834e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.6963607072830200e+08, + "cpu_time": 3.1610924250000495e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1648393510592480e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6748745366930962e+08, + "cpu_time": 3.1125045724999720e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2054395731635399e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6840401738882065e+08, + "cpu_time": 3.1316898389999890e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1881105931652184e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time_median", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6787138953804970e+08, + "cpu_time": 3.1163635574998242e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1981084655185566e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1422336012436764e+06, + "cpu_time": 2.7275539467166741e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1580306267483786e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.7826980552751501e-03, + "cpu_time": 8.7095277212625756e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7689955027747132e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.2117459177970886e+08, + "cpu_time": 5.6111646300001895e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3431717560537014e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.3171749114990234e+08, + "cpu_time": 5.7348371149998915e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2369165107268300e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.2143411040306091e+08, + "cpu_time": 5.6251707000001264e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3404725548685112e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.2259628176689148e+08, + "cpu_time": 5.6433594500003886e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3284383134207582e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.2488909363746643e+08, + "cpu_time": 5.6741179699997699e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3049488118494825e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2436231374740607e+08, + "cpu_time": 5.6577299730000734e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3107895893838573e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time_median", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2259628176689148e+08, + "cpu_time": 5.6433594500003886e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3284383134207582e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3655167735347096e+06, + "cpu_time": 4.9114730863909852e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3967337452882908e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3458766905129160e-02, + "cpu_time": 8.6809959291617167e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3280015617381862e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 310, + "real_time": 2.2542062658636319e+06, + "cpu_time": 2.5662193741936791e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8170475621629681e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 310, + "real_time": 2.2524039568980373e+06, + "cpu_time": 2.5651743161288761e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8185015114433218e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 310, + "real_time": 2.2579002134021251e+06, + "cpu_time": 2.5772364193548746e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8140748540115023e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 310, + "real_time": 2.2581288072791309e+06, + "cpu_time": 2.5720622935483009e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8138912124040260e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 310, + "real_time": 2.2556632727144226e+06, + "cpu_time": 2.5754475451615308e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8158738715778934e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2556605032314700e+06, + "cpu_time": 2.5712279896774525e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8158778023199425e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time_median", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2556632727144226e+06, + "cpu_time": 2.5720622935483009e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8158738715778934e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4407342403665161e+03, + "cpu_time": 5.3929382323329428e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9652840251318578e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0820485781747337e-03, + "cpu_time": 2.0974173639924710e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0822776855474723e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 310, + "real_time": 2.2555956466784398e+06, + "cpu_time": 2.5697965258066156e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6318566282318505e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 310, + "real_time": 2.2580808103685416e+06, + "cpu_time": 2.5740049032257898e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6278595355774635e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 310, + "real_time": 2.2583234093842967e+06, + "cpu_time": 2.5735955903226770e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6274698149781148e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 310, + "real_time": 2.2573752614158778e+06, + "cpu_time": 2.5790608387098708e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6289934332237649e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 310, + "real_time": 2.2624410846601091e+06, + "cpu_time": 2.5826470451614759e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6208677678034212e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2583632425014530e+06, + "cpu_time": 2.5758209806452859e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6274094359629229e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time_median", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2580808103685416e+06, + "cpu_time": 2.5740049032257898e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6278595355774635e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5172328320825441e+03, + "cpu_time": 5.0407438050938017e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0399945545726373e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1146270824415104e-03, + "cpu_time": 1.9569464815179087e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1137409839979066e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2650747333915490e+06, + "cpu_time": 2.5897314433657215e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2333154215480816e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2700352829347537e+06, + "cpu_time": 2.5951078349513868e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2175089626000831e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2702750065710554e+06, + "cpu_time": 2.5891324207120826e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2167468489845321e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2651855464614807e+06, + "cpu_time": 2.5894911229772754e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2329615671413643e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2708792503151596e+06, + "cpu_time": 2.5956130938510294e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2148265909454133e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2682899639347992e+06, + "cpu_time": 2.5918151831714986e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2230718782438952e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time_median", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2700352829347537e+06, + "cpu_time": 2.5897314433657215e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2175089626000831e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9011227108117264e+03, + "cpu_time": 3.2483119865188555e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.2421982523183433e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2789911153065954e-03, + "cpu_time": 1.2532961484329403e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2795384578902110e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 307, + "real_time": 2.2843666670461721e+06, + "cpu_time": 2.6150790032573831e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4344457250538968e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 307, + "real_time": 2.2898692112865006e+06, + "cpu_time": 2.6231436058634487e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4309987591645110e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 307, + "real_time": 2.2889936527886493e+06, + "cpu_time": 2.6141835211726977e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4315461277089695e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 307, + "real_time": 2.2873092125737220e+06, + "cpu_time": 2.6136961368077416e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4326003594034778e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 307, + "real_time": 2.2856492772152349e+06, + "cpu_time": 2.6111942377848574e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4336407744902808e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2872376041820562e+06, + "cpu_time": 2.6154593009772259e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4326463491642274e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time_median", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2873092125737220e+06, + "cpu_time": 2.6141835211726977e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4326003594034778e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2792132970416860e+03, + "cpu_time": 4.5310124136494533e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4277421365897100e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.9649170373655206e-04, + "cpu_time": 1.7323964521093905e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9657681564094421e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 302, + "real_time": 2.3286334718485046e+06, + "cpu_time": 2.6690772251655795e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8143544612015098e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 302, + "real_time": 2.3348544198489250e+06, + "cpu_time": 2.6710543245033007e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8068559411186099e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 302, + "real_time": 2.3285147929810826e+06, + "cpu_time": 2.6699889569536536e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8144979021626696e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 302, + "real_time": 2.3273169016773929e+06, + "cpu_time": 2.6677696721853348e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8159465499848995e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 302, + "real_time": 2.3203982373341820e+06, + "cpu_time": 2.6548191887416774e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8243427764059950e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3279435647380175e+06, + "cpu_time": 2.6665418735099100e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8151995261747368e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time_median", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3285147929810830e+06, + "cpu_time": 2.6690772251655799e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8144979021626696e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1450671547683532e+03, + "cpu_time": 6.6633594007637075e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2245259262252257e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2101339709012101e-03, + "cpu_time": 2.4988767163040553e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2110425454223651e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 291, + "real_time": 2.3903348515149448e+06, + "cpu_time": 2.7454048384882072e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4834158451452635e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 291, + "real_time": 2.3910529180221846e+06, + "cpu_time": 2.7470288591066208e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4817690989632830e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 291, + "real_time": 2.3902402918696199e+06, + "cpu_time": 2.7447946048107226e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4836327730663814e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 291, + "real_time": 2.3882213229460712e+06, + "cpu_time": 2.7434619999996847e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4882685595617957e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 291, + "real_time": 2.3894436078152503e+06, + "cpu_time": 2.7441818281786288e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4854611161902919e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3898585984336147e+06, + "cpu_time": 2.7449744261167729e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4845094785854034e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time_median", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3902402918696203e+06, + "cpu_time": 2.7447946048107226e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4836327730663814e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0784009447605856e+03, + "cpu_time": 1.3559064934096741e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4754106250276942e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.5124048153618890e-04, + "cpu_time": 4.9395960869764142e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5134585594082453e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 279, + "real_time": 2.5033355118059237e+06, + "cpu_time": 2.8849856881719963e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0471788490344529e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 279, + "real_time": 2.5041040248407791e+06, + "cpu_time": 2.8931074086018582e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0468574683780086e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 279, + "real_time": 2.5085277828691681e+06, + "cpu_time": 2.8895368817203757e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0450113480512011e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 279, + "real_time": 2.5037890258428766e+06, + "cpu_time": 2.8852342150536897e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0469891723874448e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 279, + "real_time": 2.5090032580932835e+06, + "cpu_time": 2.8945821756269815e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0448133104427145e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5057519206904066e+06, + "cpu_time": 2.8894892738349801e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0461700296587646e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time_median", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5041040248407791e+06, + "cpu_time": 2.8895368817203757e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0468574683780086e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7696639937469104e+03, + "cpu_time": 4.3993926688962774e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1559101859573692e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1053225065407867e-03, + "cpu_time": 1.5225502682200051e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1048970561070262e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 269, + "real_time": 2.5996446711630961e+06, + "cpu_time": 3.0533700557619249e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0167679291548353e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 269, + "real_time": 2.6047260181971421e+06, + "cpu_time": 3.0620893531597992e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0128335814869514e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 269, + "real_time": 2.6013029766957997e+06, + "cpu_time": 3.0580764609667012e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0154822590713969e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 269, + "real_time": 2.5986442091393536e+06, + "cpu_time": 3.0529044200744415e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0175443723927072e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 269, + "real_time": 2.6049805877792570e+06, + "cpu_time": 3.0620703754645810e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0126368789832518e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6018596925949301e+06, + "cpu_time": 3.0577021330854902e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0150530042178285e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time_median", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6013029766958002e+06, + "cpu_time": 3.0580764609667012e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0154822590713969e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8944513985514832e+03, + "cpu_time": 4.4792384085634339e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2414080987183034e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1124548363577362e-03, + "cpu_time": 1.4649034515482673e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1123320796161081e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 252, + "real_time": 2.7776860792766372e+06, + "cpu_time": 3.3640502896824433e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7749982182042307e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 252, + "real_time": 2.7794657748133419e+06, + "cpu_time": 3.3698489325393848e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7725810819542050e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 252, + "real_time": 2.7752429097225624e+06, + "cpu_time": 3.3602920753965224e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7783215167454469e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 252, + "real_time": 2.7819292237316924e+06, + "cpu_time": 3.3739850476191379e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7692403927999127e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 252, + "real_time": 2.7785254670085297e+06, + "cpu_time": 3.3643051230159118e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7738577977798361e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7785698909105523e+06, + "cpu_time": 3.3664962936506802e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7737998014967263e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time_median", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7785254670085297e+06, + "cpu_time": 3.3643051230159113e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.7738577977798361e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4464068182008677e+03, + "cpu_time": 5.3990302791871054e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3226104180297756e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.8045538325442910e-04, + "cpu_time": 1.6037535194587471e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8044162191963540e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 226, + "real_time": 3.0930701960537550e+06, + "cpu_time": 3.9610048539821319e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7801629677710462e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 226, + "real_time": 3.0942843872559280e+06, + "cpu_time": 3.9645916283185030e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7775024449507546e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 226, + "real_time": 3.0929356323044118e+06, + "cpu_time": 3.9611141238935441e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7804579510033417e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 226, + "real_time": 3.0938192452015602e+06, + "cpu_time": 3.9621902079645740e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7785214124989128e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 226, + "real_time": 3.0984188763803877e+06, + "cpu_time": 3.9676343761062422e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7684586354247868e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0945056674392084e+06, + "cpu_time": 3.9633070380529994e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7770206823297691e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time_median", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0938192452015602e+06, + "cpu_time": 3.9621902079645745e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7785214124989128e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2560437190032617e+03, + "cpu_time": 2.8165697320641953e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9366412452192634e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.2904817811182156e-04, + "cpu_time": 7.1066150187744565e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.2843827348659804e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 187, + "real_time": 3.7380471691130637e+06, + "cpu_time": 5.2039669625668572e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1220575370629134e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 187, + "real_time": 3.7495851153159207e+06, + "cpu_time": 5.2142393957216060e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1186048245357966e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 187, + "real_time": 3.7443684625693341e+06, + "cpu_time": 5.2056913475936316e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1201632643604541e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 187, + "real_time": 3.7478037529629981e+06, + "cpu_time": 5.2105477967916885e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1191365067298417e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 187, + "real_time": 3.7528065986991087e+06, + "cpu_time": 5.2214137754010959e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1176445920378454e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7465222197320857e+06, + "cpu_time": 5.2111718556149760e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1195213449453704e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time_median", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7478037529629976e+06, + "cpu_time": 5.2105477967916895e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1191365067298417e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.6343195836008726e+03, + "cpu_time": 7.0114660495253938e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6848325272263710e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5038799326816178e-03, + "cpu_time": 1.3454682063441493e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5049579312028090e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 135, + "real_time": 5.1633274907039274e+06, + "cpu_time": 7.9411197629623432e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6246515478831973e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 135, + "real_time": 5.1977916034283461e+06, + "cpu_time": 7.9775319481482422e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6138792471916463e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 135, + "real_time": 5.2420667407137379e+06, + "cpu_time": 8.0376934222217472e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6002482255420203e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 135, + "real_time": 5.2318597867808957e+06, + "cpu_time": 8.0225363407407710e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6033701861038244e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 135, + "real_time": 5.2229542699125074e+06, + "cpu_time": 8.0157645333328126e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6061040488758714e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2115999783078823e+06, + "cpu_time": 7.9989292014811831e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6096506511193120e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time_median", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2229542699125074e+06, + "cpu_time": 8.0157645333328117e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6061040488758714e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1577316482743008e+04, + "cpu_time": 3.9189168989232377e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.7894862413918339e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.0590445571756316e-03, + "cpu_time": 4.8993018943055038e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0817458959709437e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 83, + "real_time": 8.3963883681649193e+06, + "cpu_time": 1.3894815180722628e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.9981467345663958e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 83, + "real_time": 8.3882264590945588e+06, + "cpu_time": 1.3897941662650231e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0000909705781794e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 83, + "real_time": 8.3842571173985321e+06, + "cpu_time": 1.3892157819276942e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0010378695549400e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 83, + "real_time": 8.3867458020707211e+06, + "cpu_time": 1.3891749939759107e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0004440811664565e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 83, + "real_time": 8.3852268373930315e+06, + "cpu_time": 1.3884807626506796e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0008064570398719e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.3881689168243529e+06, + "cpu_time": 1.3892294445783142e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0001052225811689e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time_median", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.3867458020707201e+06, + "cpu_time": 1.3892157819276944e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0004440811664565e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8359578458797387e+03, + "cpu_time": 4.8634550365761706e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1524174268727456e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.7652127583889511e-04, + "cpu_time": 3.5008292226720123e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7617839994714474e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49, + "real_time": 1.4495956977563245e+07, + "cpu_time": 2.5256759081631869e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3147441767339230e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49, + "real_time": 1.4498311304012123e+07, + "cpu_time": 2.5366690122449636e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3143682941001873e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49, + "real_time": 1.4465938835423820e+07, + "cpu_time": 2.5235227265305150e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3195474819672794e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49, + "real_time": 1.4560329480742922e+07, + "cpu_time": 2.5339593979593903e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3045104881986456e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49, + "real_time": 1.4462255682720214e+07, + "cpu_time": 2.5264583142856915e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3201382091516671e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4496558456092468e+07, + "cpu_time": 2.5292570718367495e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3146617300303402e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time_median", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4495956977563247e+07, + "cpu_time": 2.5264583142856911e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.3147441767339230e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9320539377737747e+04, + "cpu_time": 5.7137821262611003e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2650420667829514e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7124051199346913e-03, + "cpu_time": 2.2590752794107026e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7066771725218052e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 26, + "real_time": 2.6530709690772571e+07, + "cpu_time": 4.8406277153844714e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5294786600955715e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 26, + "real_time": 2.6494082278357103e+07, + "cpu_time": 4.8389187423074767e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5329756016807170e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 26, + "real_time": 2.6550679181057673e+07, + "cpu_time": 4.8444926500001028e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5275761701748924e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 26, + "real_time": 2.6658691394214448e+07, + "cpu_time": 4.7656928269230567e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5173352662975864e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 26, + "real_time": 2.6465848231544860e+07, + "cpu_time": 4.8301825923077174e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5356778068428731e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6540002155189335e+07, + "cpu_time": 4.8239829053845651e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5286087010183282e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time_median", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6530709690772571e+07, + "cpu_time": 4.8389187423074767e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5294786600955715e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.3984666851249334e+04, + "cpu_time": 3.3003590386925801e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.0348304474237328e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7876661960550444e-03, + "cpu_time": 6.8415645399752456e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7820953256194390e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 14, + "real_time": 4.9641674384474754e+07, + "cpu_time": 9.0888550357143626e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.7037308806404018e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 14, + "real_time": 4.9753000693661824e+07, + "cpu_time": 9.0975492642850712e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.6976810670456381e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 14, + "real_time": 4.9677434776510514e+07, + "cpu_time": 9.0872035142857447e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.7017845950343542e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 14, + "real_time": 4.9670429368104257e+07, + "cpu_time": 9.0821018071429595e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.7021656488072882e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 14, + "real_time": 4.9700405714767322e+07, + "cpu_time": 9.0955455142859265e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.7005358622277870e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9688588987503737e+07, + "cpu_time": 9.0902510271428138e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.7011796107510943e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time_median", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9677434776510507e+07, + "cpu_time": 9.0888550357143626e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.7017845950343542e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1664346594588947e+04, + "cpu_time": 6.3036055699317876e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2639277421684642e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.3850935282278149e-04, + "cpu_time": 6.9344680923658650e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.3812558526567317e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 7, + "real_time": 9.5877393015793398e+07, + "cpu_time": 1.7413566442857179e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.7997784207146935e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 7, + "real_time": 9.5755475972379953e+07, + "cpu_time": 1.7377157457143050e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8033431328504748e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 7, + "real_time": 9.5719883484499797e+07, + "cpu_time": 1.7374188171429849e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8043855281485853e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 7, + "real_time": 9.6875709082399100e+07, + "cpu_time": 1.7461989528572175e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.7709263606182036e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 7, + "real_time": 9.5787452799933299e+07, + "cpu_time": 1.7393103642856595e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8024072898218555e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6003182871001109e+07, + "cpu_time": 1.7404001048571768e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.7961681464307628e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time_median", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5787452799933299e+07, + "cpu_time": 1.7393103642856592e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8024072898218555e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9124377624917607e+05, + "cpu_time": 3.5994097003428615e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4213677691643355e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.1169530171646218e-03, + "cpu_time": 2.0681507029892080e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0832700135672280e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.8634339049458504e+08, + "cpu_time": 3.2994862374999911e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8810837377975097e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.8781987950205803e+08, + "cpu_time": 3.3374029550000727e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8584349719706707e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.8586495146155357e+08, + "cpu_time": 3.3072264525000608e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8884999984036932e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.8780880421400070e+08, + "cpu_time": 3.3383646350000620e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8586035369686761e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.8630335479974747e+08, + "cpu_time": 3.3064548274998629e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8817028688349080e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8682807609438899e+08, + "cpu_time": 3.3177870215000099e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8736650227950916e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time_median", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8634339049458504e+08, + "cpu_time": 3.3072264525000608e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.8810837377975097e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1969604627545679e+05, + "cpu_time": 1.8594931822327909e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4129079994462484e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.9226864907114358e-03, + "cpu_time": 5.6046188926017690e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9167456479390663e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.5990819334983826e+08, + "cpu_time": 6.0264872150003159e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.9833769940221968e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.6190642416477203e+08, + "cpu_time": 6.0553839450000167e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.9669045706443086e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.6100058257579803e+08, + "cpu_time": 6.0373644099996686e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.9743492831470714e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.6076891422271729e+08, + "cpu_time": 6.0357428699995804e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.9762592664431605e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.6004149913787842e+08, + "cpu_time": 6.0358780249998748e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.9822723951852255e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6072512269020087e+08, + "cpu_time": 6.0381712929998922e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.9766325018883929e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time_median", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6076891422271729e+08, + "cpu_time": 6.0358780249998736e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.9762592664431605e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0746265068035375e+05, + "cpu_time": 1.0542775299932188e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6573570082067857e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2384430689453851e-03, + "cpu_time": 1.7460212352959459e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2365397824499061e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 310, + "real_time": 2.2539294572667247e+06, + "cpu_time": 2.5766783064513691e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8172707166120012e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 310, + "real_time": 2.2537848476560847e+06, + "cpu_time": 2.5773158580642710e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8173873181638443e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 310, + "real_time": 2.2538296998508517e+06, + "cpu_time": 2.5789157516126782e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8173511513629691e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 310, + "real_time": 2.2597187296337178e+06, + "cpu_time": 2.5789198322579595e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8126149711844574e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 310, + "real_time": 2.2603380801757015e+06, + "cpu_time": 2.5797348419356211e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8121183003215201e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2563201629166161e+06, + "cpu_time": 2.5783129180643796e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8153484915289585e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time_median", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2539294572667247e+06, + "cpu_time": 2.5789157516126786e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8172707166120012e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3926250192754733e+03, + "cpu_time": 1.2668567787203197e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7280340189896183e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5036097602788877e-03, + "cpu_time": 4.9135105744704897e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5027605067123833e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 311, + "real_time": 2.2597358288045768e+06, + "cpu_time": 2.5797616495177587e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6252025106552616e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 311, + "real_time": 2.2543588513957630e+06, + "cpu_time": 2.5809644147911118e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6338491518011908e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 311, + "real_time": 2.2595829069853979e+06, + "cpu_time": 2.5817875723474291e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6254478535285448e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 311, + "real_time": 2.2549498818778819e+06, + "cpu_time": 2.5805892797427424e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6328967068562289e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 311, + "real_time": 2.2536729053681018e+06, + "cpu_time": 2.5797302733118655e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6349551793817068e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2564600748863439e+06, + "cpu_time": 2.5805666379421814e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6304702804445866e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time_median", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2549498818778815e+06, + "cpu_time": 2.5805892797427420e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6328967068562289e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9557901250926093e+03, + "cpu_time": 8.6558247018148870e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7537565655358294e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3099235204688876e-03, + "cpu_time": 3.3542341339099429e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3094051729721598e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2698599509111210e+06, + "cpu_time": 2.5937835663431915e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2180664685605243e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2659562204129584e+06, + "cpu_time": 2.5970935016181529e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2305015659190910e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2637328113396093e+06, + "cpu_time": 2.5873329320386695e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2376032709904658e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2658765011596642e+06, + "cpu_time": 2.6023422038835343e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2307559532104917e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2637255649567614e+06, + "cpu_time": 2.5882222686082604e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2376264391893921e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2658302097560233e+06, + "cpu_time": 2.5937548944983617e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2309107395739928e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time_median", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2658765011596638e+06, + "cpu_time": 2.5937835663431911e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2307559532104917e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5042686157159856e+03, + "cpu_time": 6.2598033738874838e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.9852563811474611e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1052322477356487e-03, + "cpu_time": 2.4134136140485794e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1043223556121383e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 308, + "real_time": 2.2760446806281032e+06, + "cpu_time": 2.6086371720777736e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4396905420572525e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 308, + "real_time": 2.2759771544498377e+06, + "cpu_time": 2.6031034188312711e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4397332563700916e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 308, + "real_time": 2.2870047296119202e+06, + "cpu_time": 2.6140086948051797e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4327910902729254e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 308, + "real_time": 2.2756134797879434e+06, + "cpu_time": 2.6022708474022825e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4399633457547253e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 308, + "real_time": 2.2814608067153143e+06, + "cpu_time": 2.6147448571429639e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4362727557514803e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2792201702386239e+06, + "cpu_time": 2.6085529980518939e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4376901980412953e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time_median", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2760446806281027e+06, + "cpu_time": 2.6086371720777731e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4396905420572525e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9806856489046231e+03, + "cpu_time": 5.8582028344196569e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1366010111464606e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1852586748489365e-03, + "cpu_time": 2.2457672275758437e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1816946484157409e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 303, + "real_time": 2.3085359390210300e+06, + "cpu_time": 2.6448457194716623e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8388555227687530e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 303, + "real_time": 2.3144448833792121e+06, + "cpu_time": 2.6572511188118877e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8316077203063037e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 303, + "real_time": 2.3077396397777023e+06, + "cpu_time": 2.6438287590759117e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8398350866960403e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 303, + "real_time": 2.3155923342645760e+06, + "cpu_time": 2.6516069471947602e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8302045671097804e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 303, + "real_time": 2.3163453562080665e+06, + "cpu_time": 2.6524683201318947e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8292844944021903e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3125316305301175e+06, + "cpu_time": 2.6500001729372228e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8339574782566138e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time_median", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3144448833792121e+06, + "cpu_time": 2.6516069471947597e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8316077203063037e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0774279076443017e+03, + "cpu_time": 5.6103571576357208e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9994900644653251e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7631879511674424e-03, + "cpu_time": 2.1171157703802259e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7641372895760236e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 296, + "real_time": 2.3809703282470740e+06, + "cpu_time": 2.7350888277024948e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5049825041918211e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 296, + "real_time": 2.3710621855964228e+06, + "cpu_time": 2.7260726689191288e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5279866043255977e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 296, + "real_time": 2.3715016432106495e+06, + "cpu_time": 2.7260868412159542e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5269622256111369e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 296, + "real_time": 2.3704010905144182e+06, + "cpu_time": 2.7248198277027006e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5295283369767226e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 296, + "real_time": 2.3699789205754832e+06, + "cpu_time": 2.7242333817565776e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5305133249106213e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3727828336288100e+06, + "cpu_time": 2.7272603094593715e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5239945992031805e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time_median", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3710621855964232e+06, + "cpu_time": 2.7260726689191284e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5279866043255977e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6144277904235942e+03, + "cpu_time": 4.4494791317864647e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0715860986173719e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9447324571910060e-03, + "cpu_time": 1.6314831101210472e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9398753553668299e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 278, + "real_time": 2.4982848097469737e+06, + "cpu_time": 2.8915120899283602e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0492958968379188e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 278, + "real_time": 2.4973478355853679e+06, + "cpu_time": 2.8942312517988645e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0496895797399186e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 278, + "real_time": 2.4973230801603133e+06, + "cpu_time": 2.8904248345325184e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0496999850863186e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 278, + "real_time": 2.5029207526190248e+06, + "cpu_time": 2.8922960791365448e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0473523771205933e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 278, + "real_time": 2.5057041114732195e+06, + "cpu_time": 2.8952160827337485e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0461889686004204e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5003161179169798e+06, + "cpu_time": 2.8927360676260078e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0484453614770339e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time_median", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4982848097469737e+06, + "cpu_time": 2.8922960791365444e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0492958968379188e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7983478112315920e+03, + "cpu_time": 1.9633873999191462e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5915003537542804e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5191470326544171e-03, + "cpu_time": 6.7873022426496261e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5179621296737855e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 271, + "real_time": 2.5867501418089075e+06, + "cpu_time": 3.0472185608857460e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0268211897472510e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 271, + "real_time": 2.5946887259842395e+06, + "cpu_time": 3.0538424169743448e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0206200256299436e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 271, + "real_time": 2.5884398814235777e+06, + "cpu_time": 3.0467057453874131e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0254980761293733e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 271, + "real_time": 2.5909798231828148e+06, + "cpu_time": 3.0485265756455716e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0235124770518416e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 271, + "real_time": 2.5872608579736464e+06, + "cpu_time": 3.0479483025831333e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0264211023956221e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5896238860746371e+06, + "cpu_time": 3.0488483202952417e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0245745741908064e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time_median", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5884398814235777e+06, + "cpu_time": 3.0479483025831329e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0254980761293733e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2686137609130424e+03, + "cpu_time": 2.8766862930992233e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5532483402520797e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2621963283894562e-03, + "cpu_time": 9.4353211143696812e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2611283243407205e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 254, + "real_time": 2.7484369480718425e+06, + "cpu_time": 3.3379927913386840e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8151721135011858e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 254, + "real_time": 2.7476782006702907e+06, + "cpu_time": 3.3416150905511105e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8162256400483942e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 254, + "real_time": 2.7477327453735191e+06, + "cpu_time": 3.3514589803149379e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8161498849025059e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 254, + "real_time": 2.7450979778939108e+06, + "cpu_time": 3.3404015629919628e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8198126567580169e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 254, + "real_time": 2.7455981476159077e+06, + "cpu_time": 3.3363370433069286e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8191167957718527e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7469088039250942e+06, + "cpu_time": 3.3415610937007247e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8172954181963915e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time_median", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7476782006702907e+06, + "cpu_time": 3.3404015629919628e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8162256400483942e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4665423853159941e+03, + "cpu_time": 5.9020985848628243e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0383204409513241e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3388826859502312e-04, + "cpu_time": 1.7662698419577136e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3396979212952725e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 227, + "real_time": 3.0794551766309314e+06, + "cpu_time": 3.9961192907490754e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8101397153453064e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 227, + "real_time": 3.0872818310273243e+06, + "cpu_time": 4.0142229691626616e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7928751399484360e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 227, + "real_time": 3.0885970851792111e+06, + "cpu_time": 4.0014680837004506e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.7899824488706851e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 227, + "real_time": 3.0796941572857718e+06, + "cpu_time": 4.0034443392072124e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8096112564901054e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 227, + "real_time": 3.0712782464806596e+06, + "cpu_time": 3.9811134449341265e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8282709402936733e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0812612993207802e+06, + "cpu_time": 3.9992736255507059e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8061759001896417e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time_median", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0796941572857718e+06, + "cpu_time": 4.0014680837004511e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8096112564901054e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9899913004675373e+03, + "cpu_time": 1.2097885780792076e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5450262046968653e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2685486953048675e-03, + "cpu_time": 3.0250207696469328e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2700356666565374e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 194, + "real_time": 3.6101138780911253e+06, + "cpu_time": 5.0826293659795113e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1618204138806198e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 194, + "real_time": 3.6025072601575828e+06, + "cpu_time": 5.0762658969074208e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1642735731271033e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 194, + "real_time": 3.5979348959727692e+06, + "cpu_time": 5.0626972010310553e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1657531671000376e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 194, + "real_time": 3.5925707719816836e+06, + "cpu_time": 5.0568943659799490e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1674937715106993e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 194, + "real_time": 3.5909823029177245e+06, + "cpu_time": 5.0539988814431671e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1680102117440312e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5988218218241772e+06, + "cpu_time": 5.0664971422682200e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1654702274724982e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time_median", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5979348959727688e+06, + "cpu_time": 5.0626972010310553e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1657531671000376e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7848138821449102e+03, + "cpu_time": 1.2435198013726003e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5188998732621348e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1631562404495286e-03, + "cpu_time": 2.4543975185504375e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1612734618925078e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 144, + "real_time": 4.8496745390972747e+06, + "cpu_time": 7.5259406597222919e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7297259707579198e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 144, + "real_time": 4.8983302905172519e+06, + "cpu_time": 7.5845295347218346e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7125443778749723e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 144, + "real_time": 4.8988481123362565e+06, + "cpu_time": 7.5883428055555485e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7123633571891823e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 144, + "real_time": 4.8838624788913876e+06, + "cpu_time": 7.5735561736109955e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7176175693432245e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 144, + "real_time": 4.9532038230811143e+06, + "cpu_time": 7.6718353819439746e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6935721403004794e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8967838487846572e+06, + "cpu_time": 7.5888409111109301e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7131646830931559e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time_median", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8983302905172529e+06, + "cpu_time": 7.5845295347218364e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7125443778749723e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7336694500200472e+04, + "cpu_time": 5.2671109466045338e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3028805393465664e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.6247381246911982e-03, + "cpu_time": 6.9406000314130734e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6051097259032183e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 88, + "real_time": 7.8122451933185486e+06, + "cpu_time": 1.3323577829545094e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1475536910117435e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 88, + "real_time": 7.8531108500266615e+06, + "cpu_time": 1.3388211204546576e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1363783499812741e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 88, + "real_time": 7.8488417784683406e+06, + "cpu_time": 1.3386185488636732e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1375403497143734e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 88, + "real_time": 7.8318580688739363e+06, + "cpu_time": 1.3347926465908958e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1421756947661624e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 88, + "real_time": 7.8284707745875828e+06, + "cpu_time": 1.3346577420454703e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1431025909250908e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8349053330550147e+06, + "cpu_time": 1.3358495681818414e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1413501352797289e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time_median", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8318580688739363e+06, + "cpu_time": 1.3347926465908958e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1421756947661624e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6506337066691474e+04, + "cpu_time": 2.7940773874289280e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5130330179160004e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1067691778038707e-03, + "cpu_time": 2.0916108025784731e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1075642621735252e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.3725475097695986e+07, + "cpu_time": 2.5012464254901338e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4446827349264274e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.3694337934401689e+07, + "cpu_time": 2.5028846745098077e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4502412720302134e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.3714952704807123e+07, + "cpu_time": 2.5021164254903670e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4465583456397262e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3715244537475063e+07, + "cpu_time": 2.5066578058822773e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4465062878257127e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.3740430549517566e+07, + "cpu_time": 2.5062390823530436e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4420218769038582e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3718088164779488e+07, + "cpu_time": 2.5038288827451263e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4460021034651875e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time_median", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3715244537475064e+07, + "cpu_time": 2.5028846745098077e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4465062878257127e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6850368705283232e+04, + "cpu_time": 2.4650047047888635e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0048749630691791e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2283321482468466e-03, + "cpu_time": 9.8449407696196201e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2284842105459561e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27, + "real_time": 2.5511406638004161e+07, + "cpu_time": 4.8405947111110367e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6305434644292955e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27, + "real_time": 2.5325498233238857e+07, + "cpu_time": 4.8206645592594445e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6498536527080798e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27, + "real_time": 2.5446310294447120e+07, + "cpu_time": 4.8338933629633240e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6372728786005750e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27, + "real_time": 2.5477717803032309e+07, + "cpu_time": 4.8418816370367825e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6340217957831697e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27, + "real_time": 2.5425149196827855e+07, + "cpu_time": 4.8284418111111164e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6394678544648533e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5437216433110062e+07, + "cpu_time": 4.8330952162963405e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6382319291971946e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time_median", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5446310294447120e+07, + "cpu_time": 4.8338933629633240e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6372728786005750e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0451792304878967e+04, + "cpu_time": 8.8070435472368932e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3203391405589944e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7696345034504723e-03, + "cpu_time": 1.8222367143815217e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7747140270516510e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15, + "real_time": 4.6138092378775276e+07, + "cpu_time": 8.8417745333337441e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9090437224436188e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15, + "real_time": 4.6338262408971786e+07, + "cpu_time": 8.8564536533332720e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8964773606619620e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15, + "real_time": 4.6274092296759285e+07, + "cpu_time": 8.8422008599998966e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9004940202662749e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15, + "real_time": 4.6637549748023354e+07, + "cpu_time": 8.9007679399999991e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8778897846297894e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15, + "real_time": 4.6032821138699852e+07, + "cpu_time": 8.8151940933335036e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9156963375239019e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6284163594245911e+07, + "cpu_time": 8.8512782160000831e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.8999202451051097e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time_median", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6274092296759285e+07, + "cpu_time": 8.8422008599998966e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9004940202662749e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3044506743356731e+05, + "cpu_time": 3.1427695660195395e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4401794946486358e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.9789182635724769e-03, + "cpu_time": 3.5506392289629848e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9662727693272658e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.7771806865930557e+07, + "cpu_time": 1.6733663362499839e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0583334852617202e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.8093709200620651e+07, + "cpu_time": 1.6769270125000447e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0471580597052302e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.7507283315062523e+07, + "cpu_time": 1.6546858362499961e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0675784441110005e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.7659886106848717e+07, + "cpu_time": 1.6713385725000763e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0622382474100385e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.7684210389852524e+07, + "cpu_time": 1.6728462749999550e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0613887586660113e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7743379175662994e+07, + "cpu_time": 1.6698328065000111e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0593393990308003e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time_median", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7684210389852524e+07, + "cpu_time": 1.6728462749999553e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0613887586660113e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1780298041485302e+05, + "cpu_time": 8.7119547557947820e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5809292130819950e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4822725368122598e-03, + "cpu_time": 5.2172617054129745e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4779627966363053e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6998168081045151e+08, + "cpu_time": 3.1664831074999708e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1584045377141013e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6995543614029884e+08, + "cpu_time": 3.1375808149999785e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1588922613620377e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.7029255628585815e+08, + "cpu_time": 3.1664943249998599e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1526387512722077e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.7059879377484322e+08, + "cpu_time": 3.1513186224998432e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1469795308668113e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.7028475552797318e+08, + "cpu_time": 3.1681567724999124e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1527831738984213e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7022264450788498e+08, + "cpu_time": 3.1580067284999138e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1539396510227160e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time_median", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7028475552797318e+08, + "cpu_time": 3.1664831074999708e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1527831738984213e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6443218493749539e+05, + "cpu_time": 1.3312216125254431e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8973619026083825e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5534489297940998e-03, + "cpu_time": 4.2153856117899636e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5527760339422898e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.2639713585376740e+08, + "cpu_time": 5.6838573800001764e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2896790628734512e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.2704120874404907e+08, + "cpu_time": 5.6875384349996233e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2832003897109437e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.2643881440162659e+08, + "cpu_time": 5.6811588199997234e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2892590483402081e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.2616929709911346e+08, + "cpu_time": 5.6786998600000513e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2919769995203466e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.2659913599491119e+08, + "cpu_time": 5.6668310050002897e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2876444107210689e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2652911841869360e+08, + "cpu_time": 5.6796170999999738e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2883519822332039e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time_median", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2643881440162659e+08, + "cpu_time": 5.6811588199997234e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2892590483402081e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2487865318262315e+05, + "cpu_time": 7.8649535565121868e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2697180776329939e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.9494542709065810e-04, + "cpu_time": 1.3847682718809025e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9433336069225929e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 313, + "real_time": 2.2467311505918591e+06, + "cpu_time": 2.5798558178914287e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8230930740961088e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 313, + "real_time": 2.2435019673488010e+06, + "cpu_time": 2.5779414376993719e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8257171420448271e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 313, + "real_time": 2.2514805193252554e+06, + "cpu_time": 2.5920550543130189e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8192473640533774e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 313, + "real_time": 2.2430798134055380e+06, + "cpu_time": 2.5770478785942397e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8260607471569548e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 313, + "real_time": 2.2431983636281551e+06, + "cpu_time": 2.5887573322684043e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8259642421346628e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2455983628599215e+06, + "cpu_time": 2.5831315041532926e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8240165138971861e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time_median", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2435019673488010e+06, + "cpu_time": 2.5798558178914287e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8257171420448271e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6187337058094154e+03, + "cpu_time": 6.8183044018375031e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9352258134889512e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6114785999401572e-03, + "cpu_time": 2.6395498606535048e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6092101091878604e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 312, + "real_time": 2.2543973598378496e+06, + "cpu_time": 2.5901749102563197e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6337870802817210e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 312, + "real_time": 2.2435978552004178e+06, + "cpu_time": 2.5915998589745653e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6512782275182814e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 312, + "real_time": 2.2436537166448454e+06, + "cpu_time": 2.5793571602566070e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6511873196949027e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 312, + "real_time": 2.2560974100950914e+06, + "cpu_time": 2.5916415128203854e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6310488914814712e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 312, + "real_time": 2.2434188524643197e+06, + "cpu_time": 2.5793039198716977e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6515695635709600e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2482330388485049e+06, + "cpu_time": 2.5864154724359149e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6437742165094675e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time_median", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2436537166448450e+06, + "cpu_time": 2.5901749102563197e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6511873196949027e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4319261928004125e+03, + "cpu_time": 6.4945493549201074e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0412834664114642e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8608805589364983e-03, + "cpu_time": 2.5110232381974841e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8577057867459079e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 311, + "real_time": 2.2512832851152136e+06, + "cpu_time": 2.6009064758843714e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2776269909370905e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 311, + "real_time": 2.2519268664468522e+06, + "cpu_time": 2.5923213729905635e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2755471077313861e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 311, + "real_time": 2.2585872452933691e+06, + "cpu_time": 2.6024865787781789e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2540921472669858e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 311, + "real_time": 2.2514684798070183e+06, + "cpu_time": 2.5895642379422076e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2770283692376325e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 311, + "real_time": 2.2576395759113445e+06, + "cpu_time": 2.5958957749194168e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2571371333204275e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2541810905147595e+06, + "cpu_time": 2.5962348881029477e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2682863496987047e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time_median", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2519268664468518e+06, + "cpu_time": 2.5958957749194168e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2755471077313861e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6129067350447754e+03, + "cpu_time": 5.4962610143215079e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1642266747482265e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6027579817111056e-03, + "cpu_time": 2.1170122316388680e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6017897737291096e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 307, + "real_time": 2.2624560741022085e+06, + "cpu_time": 2.6045610716609862e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4483375113924831e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 307, + "real_time": 2.2684272260371754e+06, + "cpu_time": 2.6167442214984256e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4445250711103478e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 307, + "real_time": 2.2628282353051421e+06, + "cpu_time": 2.6047750260587242e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4480993072627645e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 307, + "real_time": 2.2624644681398761e+06, + "cpu_time": 2.6107098078178982e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4483321378717950e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 307, + "real_time": 2.2681838429323626e+06, + "cpu_time": 2.6105062052118299e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4446800730948133e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2648719693033532e+06, + "cpu_time": 2.6094592664495734e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4467948201464407e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time_median", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2628282353051421e+06, + "cpu_time": 2.6105062052118303e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4480993072627645e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1391805260308984e+03, + "cpu_time": 5.0414512983292934e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0042981617248843e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3860300134300623e-03, + "cpu_time": 1.9319908009863994e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3853368382408327e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 304, + "real_time": 2.3107426187427910e+06, + "cpu_time": 2.6615917368418784e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8361445133883525e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 304, + "real_time": 2.3204474127851427e+06, + "cpu_time": 2.6714592697371729e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8242829222895291e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 304, + "real_time": 2.3027246467168690e+06, + "cpu_time": 2.6604286710524224e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8460198267056622e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 304, + "real_time": 2.3111162843538057e+06, + "cpu_time": 2.6624383914473900e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8356859602295626e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 304, + "real_time": 2.3032431000549542e+06, + "cpu_time": 2.6609533782896139e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8453791959014811e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3096548125307122e+06, + "cpu_time": 2.6633742894736952e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8375024837029178e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time_median", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3107426187427910e+06, + "cpu_time": 2.6615917368418779e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8361445133883525e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2273213844089769e+03, + "cpu_time": 4.5814240064237547e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8669099993141630e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1291781547606789e-03, + "cpu_time": 1.7201577805007204e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1248994671338286e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 296, + "real_time": 2.3583934148395988e+06, + "cpu_time": 2.7339229560807878e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5576817326262161e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 296, + "real_time": 2.3671994119492434e+06, + "cpu_time": 2.7421733513513976e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5370071206662834e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 296, + "real_time": 2.3608226250429209e+06, + "cpu_time": 2.7299706013515964e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5519630576912589e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 296, + "real_time": 2.3641507581782504e+06, + "cpu_time": 2.7388758006755882e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5441472819187082e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 296, + "real_time": 2.3570469261235180e+06, + "cpu_time": 2.7319432635136009e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5608566188185997e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3615226272267061e+06, + "cpu_time": 2.7353771945945946e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5503311623442143e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time_median", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3608226250429214e+06, + "cpu_time": 2.7339229560807878e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5519630576912589e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1667382527417221e+03, + "cpu_time": 5.0389962788816538e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.7883816793175778e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7644286803362120e-03, + "cpu_time": 1.8421577429391688e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7635671445563616e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 280, + "real_time": 2.4897267285268754e+06, + "cpu_time": 2.8928891321429871e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0529027021174556e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 280, + "real_time": 2.4959656393288500e+06, + "cpu_time": 2.8955053500003000e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0502708685945249e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 280, + "real_time": 2.5009125196707570e+06, + "cpu_time": 2.9005259714283836e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0481934011610731e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 280, + "real_time": 2.4902827605339033e+06, + "cpu_time": 2.8936781428569704e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0526676092950895e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 280, + "real_time": 2.4952553345688754e+06, + "cpu_time": 2.8982864892855431e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0505698409629597e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4944285965258526e+06, + "cpu_time": 2.8961770171428369e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0509208844262205e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time_median", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4952553345688749e+06, + "cpu_time": 2.8955053500003000e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0505698409629597e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5930048570505869e+03, + "cpu_time": 3.1964422926262314e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9341384507061535e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8413054049522821e-03, + "cpu_time": 1.1036764236806271e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8404225088381892e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 270, + "real_time": 2.5784969855858772e+06, + "cpu_time": 3.0504236925922767e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0333085628210384e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 270, + "real_time": 2.5843920686316709e+06, + "cpu_time": 3.0566709629628924e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0286705193209669e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 270, + "real_time": 2.5766510326691247e+06, + "cpu_time": 3.0448905814814176e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0347652567328677e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 270, + "real_time": 2.5760762091657077e+06, + "cpu_time": 3.0480800074070869e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0352192925604355e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 270, + "real_time": 2.5813361060702140e+06, + "cpu_time": 3.0495282370368303e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0310721984909123e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5793904804245187e+06, + "cpu_time": 3.0499186962961005e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0326071659852445e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time_median", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5784969855858767e+06, + "cpu_time": 3.0495282370368303e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0333085628210384e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4672988316820056e+03, + "cpu_time": 4.3203964008351268e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7306515065090236e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3442318477934964e-03, + "cpu_time": 1.4165611713131653e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3434231425556464e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 254, + "real_time": 2.7343525292762504e+06, + "cpu_time": 3.3397055708662705e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8348237426340383e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 254, + "real_time": 2.7386404594083824e+06, + "cpu_time": 3.3459995669291378e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8288195020186031e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 254, + "real_time": 2.7329270429102574e+06, + "cpu_time": 3.3381841417323141e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8368239749400175e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 254, + "real_time": 2.7362996112833931e+06, + "cpu_time": 3.3402710039372034e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8320949784742016e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 254, + "real_time": 2.7387027993548927e+06, + "cpu_time": 3.3445048031494948e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8287323482014710e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7361844884466352e+06, + "cpu_time": 3.3417330173228839e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8322589092536664e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time_median", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7362996112833926e+06, + "cpu_time": 3.3402710039372034e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8320949784742016e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5668055189123779e+03, + "cpu_time": 3.3439790086191119e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5955258052195929e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.3809665603710227e-04, + "cpu_time": 1.0006721037511331e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.3822622384347795e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 228, + "real_time": 3.0801579041667935e+06, + "cpu_time": 4.0088476842106301e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8085860051622772e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 228, + "real_time": 3.0833544554760712e+06, + "cpu_time": 4.0094556140353801e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8015274607025313e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 228, + "real_time": 3.0810295301898005e+06, + "cpu_time": 4.0065427500003516e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8066598500625503e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 228, + "real_time": 3.0687950651586186e+06, + "cpu_time": 4.0057992543862388e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8337961821233678e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 228, + "real_time": 3.0677016628415962e+06, + "cpu_time": 3.9933231096490705e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8362319106917942e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0762077235665764e+06, + "cpu_time": 4.0047936824563341e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8173602817485046e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time_median", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0801579041667935e+06, + "cpu_time": 4.0065427500003511e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8085860051622772e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.3693638904654481e+03, + "cpu_time": 6.5919077091482741e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6343520048019032e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3956002171145185e-03, + "cpu_time": 1.6460043217769807e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3973384671738770e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 195, + "real_time": 3.5923625521648386e+06, + "cpu_time": 5.0736965538461823e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1675614415567319e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 195, + "real_time": 3.5915979040929903e+06, + "cpu_time": 5.0684022307689535e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1678100143727572e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 195, + "real_time": 3.5927030687721875e+06, + "cpu_time": 5.0710353846153133e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1674507800148957e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 195, + "real_time": 3.5899756750903833e+06, + "cpu_time": 5.0678629435896464e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1683377213675416e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 195, + "real_time": 3.5855482141367900e+06, + "cpu_time": 5.0643699025636707e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1697803932640092e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5904374828514382e+06, + "cpu_time": 5.0690734030767540e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1681880701151872e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time_median", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5915979040929908e+06, + "cpu_time": 5.0684022307689535e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1678100143727572e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9284622175851514e+03, + "cpu_time": 3.5093267196946304e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.5355580501615116e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.1562824351405711e-04, + "cpu_time": 6.9230142091937139e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1626908321545124e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 142, + "real_time": 4.9241766805799911e+06, + "cpu_time": 7.6808203521124255e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7035554457424450e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 142, + "real_time": 4.9255400909189610e+06, + "cpu_time": 7.6719397957749031e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7030838943866830e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 142, + "real_time": 4.9171062402794482e+06, + "cpu_time": 7.6714108521128492e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7060050342787104e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 142, + "real_time": 4.9532190929363733e+06, + "cpu_time": 7.7051032816899344e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6935669193318594e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 142, + "real_time": 4.9255288756964076e+06, + "cpu_time": 7.6838829154928112e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7030877722372417e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9291141960822362e+06, + "cpu_time": 7.6826314394365847e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7018598131953881e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time_median", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9255288756964076e+06, + "cpu_time": 7.6808203521124255e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7030877722372417e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3921609678953591e+04, + "cpu_time": 1.3693454166081146e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7915648374033300e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8243633896773545e-03, + "cpu_time": 1.7823911343435958e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8154873863592534e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 82, + "real_time": 7.7598416360049713e+06, + "cpu_time": 1.3273240365853181e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1620564938020406e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 82, + "real_time": 7.7920330109120142e+06, + "cpu_time": 1.3321902939023359e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1531243484858284e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 82, + "real_time": 7.7439938535595816e+06, + "cpu_time": 1.3247519853658702e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1664810583867178e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 82, + "real_time": 7.7931686565734260e+06, + "cpu_time": 1.3331182317072921e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1528105882642045e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 82, + "real_time": 7.7372620836263746e+06, + "cpu_time": 1.3235539548780130e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1683659954474092e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7652598481352730e+06, + "cpu_time": 1.3281877004877660e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1605676968772402e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time_median", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7598416360049713e+06, + "cpu_time": 1.3273240365853181e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1620564938020406e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6273457133225493e+04, + "cpu_time": 4.3113960690910448e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3070854403106580e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.3834614226766302e-03, + "cpu_time": 3.2460743820378096e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3820210544070880e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53, + "real_time": 1.3679108477004293e+07, + "cpu_time": 2.4964502433961589e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4529692162619925e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53, + "real_time": 1.3750236393567527e+07, + "cpu_time": 2.5022043283019111e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4402803733394022e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53, + "real_time": 1.3700772024128797e+07, + "cpu_time": 2.5005635396226197e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4490906016760507e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53, + "real_time": 1.3731368307797413e+07, + "cpu_time": 2.4998768150942609e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4436335292926326e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53, + "real_time": 1.3718975936326215e+07, + "cpu_time": 2.4960782037735630e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4458408671125269e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3716092227764850e+07, + "cpu_time": 2.4990346260377031e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4463629175365214e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time_median", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3718975936326217e+07, + "cpu_time": 2.4998768150942612e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4458408671125269e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7431755842976687e+04, + "cpu_time": 2.6698940911671270e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8939473931295993e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9999687511175999e-03, + "cpu_time": 1.0683701871711665e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0004993364017253e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 28, + "real_time": 2.5307103858462401e+07, + "cpu_time": 4.8164615428569436e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6517796890282874e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 28, + "real_time": 2.5342475556369338e+07, + "cpu_time": 4.8164641178573214e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6480784740515804e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 28, + "real_time": 2.5283840884055410e+07, + "cpu_time": 4.8163330892854057e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6542195194053941e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 28, + "real_time": 2.5401933052177940e+07, + "cpu_time": 4.8248415785715386e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6418802010914736e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 28, + "real_time": 2.5421201384493284e+07, + "cpu_time": 4.8304818999996863e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6398777534147477e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5351310947111677e+07, + "cpu_time": 4.8209164457141794e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6471671273982968e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time_median", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5342475556369338e+07, + "cpu_time": 4.8164641178573221e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6480784740515804e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.9227972664591718e+04, + "cpu_time": 6.4726563828928840e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1832247528550988e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3362883595311537e-03, + "cpu_time": 1.3426194906669892e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3357893382924141e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15, + "real_time": 4.5261677106221519e+07, + "cpu_time": 8.6318898599999249e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9653723984865527e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15, + "real_time": 4.5544216285149254e+07, + "cpu_time": 8.7257309400001764e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9469763440361314e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15, + "real_time": 4.5757796367009483e+07, + "cpu_time": 8.7694692799997628e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9332209733939128e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15, + "real_time": 4.5727897683779396e+07, + "cpu_time": 8.7646090533333629e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9351388276835155e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15, + "real_time": 4.5735599348942436e+07, + "cpu_time": 8.7725882000002772e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9346445637670116e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5605437358220421e+07, + "cpu_time": 8.7328574666667014e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9430706214734249e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time_median", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5727897683779396e+07, + "cpu_time": 8.7646090533333629e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9351388276835155e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1039906924963338e+05, + "cpu_time": 5.9523869751568360e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3634068497384043e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.6134645655735336e-03, + "cpu_time": 6.8160816753016808e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6325998424591851e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.7242315523326397e+07, + "cpu_time": 1.6735660612499714e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0768951327091627e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.6270649917423725e+07, + "cpu_time": 1.6512092562500414e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1115501767627835e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.6642671376466751e+07, + "cpu_time": 1.6571775962499657e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0981899765490203e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.6534993723034859e+07, + "cpu_time": 1.6542719225000724e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1020451316973381e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.7502876296639442e+07, + "cpu_time": 1.6743091924999475e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0677329404577446e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6838701367378250e+07, + "cpu_time": 1.6621068057499999e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0912826716352100e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time_median", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6642671376466751e+07, + "cpu_time": 1.6571775962499657e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0981899765490203e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1414522848530125e+05, + "cpu_time": 1.1007405075032634e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8275012221132986e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.9206922764789829e-03, + "cpu_time": 6.6225618215104490e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9117894292940766e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6837120056152344e+08, + "cpu_time": 3.1233410375000405e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1886148593673859e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6812176257371902e+08, + "cpu_time": 3.1334652250001228e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1933457262238116e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6830627620220184e+08, + "cpu_time": 3.1237660625001240e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1898448715899787e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.6851808503270149e+08, + "cpu_time": 3.1424648724998862e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1858355849214549e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6852339729666710e+08, + "cpu_time": 3.1310815025000238e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1857351596994996e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6836814433336258e+08, + "cpu_time": 3.1308237400000393e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1886752403604264e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time_median", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6837120056152344e+08, + "cpu_time": 3.1310815025000238e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1886148593673859e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6667833813190699e+05, + "cpu_time": 7.8797445832739270e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1580388920214395e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.8996362282101368e-04, + "cpu_time": 2.5168279142005700e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9039213904532851e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.1837236881256104e+08, + "cpu_time": 5.6323866750000203e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3725974022329688e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.2410241663455963e+08, + "cpu_time": 5.6631119849998868e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3129707428584008e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.2256808876991272e+08, + "cpu_time": 5.6769834950000584e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3287292245634942e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.1846623122692108e+08, + "cpu_time": 5.6185022400001121e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3716033874715972e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.2743707299232483e+08, + "cpu_time": 5.7101642550003362e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.2792310723629289e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2218923568725586e+08, + "cpu_time": 5.6602297300000834e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3330263658978782e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time_median", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2256808876991272e+08, + "cpu_time": 5.6631119849998868e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3287292245634942e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8656442647588127e+06, + "cpu_time": 3.6396675521337381e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9902006703254409e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1998055293539150e-02, + "cpu_time": 6.4302470495904845e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1971704488003736e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 312, + "real_time": 2.2417214021492652e+06, + "cpu_time": 2.5715869487178894e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8271672813905121e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 312, + "real_time": 2.2474860590404999e+06, + "cpu_time": 2.5833100544871697e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8224807150744556e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 312, + "real_time": 2.2410865483554797e+06, + "cpu_time": 2.5711752756409510e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8276848803566578e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 312, + "real_time": 2.2409455111930864e+06, + "cpu_time": 2.5772109262818824e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8277999083606799e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 312, + "real_time": 2.2414224199914876e+06, + "cpu_time": 2.5770434647435024e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8274110062732201e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2425323881459637e+06, + "cpu_time": 2.5760653339742785e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8265087582911053e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time_median", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2414224199914872e+06, + "cpu_time": 2.5770434647435024e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.8274110062732201e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7855526418453437e+03, + "cpu_time": 4.9679900828802001e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2651251544521842e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2421460026931105e-03, + "cpu_time": 1.9285186665726872e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2401392241729251e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 312, + "real_time": 2.2488173398857890e+06, + "cpu_time": 2.5859493621796812e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6428036438104166e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 312, + "real_time": 2.2416342249832666e+06, + "cpu_time": 2.5737435000001467e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6544766798700853e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 312, + "real_time": 2.2470891178669170e+06, + "cpu_time": 2.5783613205127208e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6456053010378061e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 312, + "real_time": 2.2482009483680418e+06, + "cpu_time": 2.5803277660255381e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6438023949534111e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 312, + "real_time": 2.2483804204966873e+06, + "cpu_time": 2.5800882147436622e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6435115362685439e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2468244103201400e+06, + "cpu_time": 2.5796940326923500e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6460399111880530e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time_median", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2482009483680413e+06, + "cpu_time": 2.5800882147436617e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.6438023949534111e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9705170776628875e+03, + "cpu_time": 4.3838084325220425e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8280569837442108e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3220957828385138e-03, + "cpu_time": 1.6993520847690578e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3241920278845769e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 311, + "real_time": 2.2497546231715530e+06, + "cpu_time": 2.5840314598070695e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2825719886300033e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 311, + "real_time": 2.2573195281064203e+06, + "cpu_time": 2.5922235948554100e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2581660664336318e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 311, + "real_time": 2.2562184704952780e+06, + "cpu_time": 2.5901607652735873e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2617081254562354e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 311, + "real_time": 2.2508376203118032e+06, + "cpu_time": 2.5911296527330931e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2790679577011708e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 311, + "real_time": 2.2572808355464335e+06, + "cpu_time": 2.5912145755627449e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2582904802954337e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2542822155262977e+06, + "cpu_time": 2.5897520096463813e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2679609237032952e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time_median", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2562184704952780e+06, + "cpu_time": 2.5911296527330931e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 7.2617081254562354e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6854579688250242e+03, + "cpu_time": 3.2801488002415549e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1889108598280980e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6348698239473072e-03, + "cpu_time": 1.2665879929906665e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6358245074635099e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 309, + "real_time": 2.2733881519534732e+06, + "cpu_time": 2.6104258122977274e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4413728677103894e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 309, + "real_time": 2.2653881136424430e+06, + "cpu_time": 2.6102105210357024e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4464629615855718e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 309, + "real_time": 2.2659550848926348e+06, + "cpu_time": 2.6102155728155612e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4461010378567414e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 309, + "real_time": 2.2714303393507660e+06, + "cpu_time": 2.6097368543690392e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4426152293697877e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 309, + "real_time": 2.2657676614989741e+06, + "cpu_time": 2.6044042006473802e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4462206587555198e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2683858702676585e+06, + "cpu_time": 2.6089985922330823e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4445545510556022e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time_median", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2659550848926348e+06, + "cpu_time": 2.6102105210357024e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.4461010378567414e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7430566900414187e+03, + "cpu_time": 2.5807089774333776e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3818923440558508e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6500969870702630e-03, + "cpu_time": 9.8915690683623928e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6488767020361350e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 305, + "real_time": 2.3038666848032200e+06, + "cpu_time": 2.6510516459017606e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8446090406310823e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 305, + "real_time": 2.2950713705943255e+06, + "cpu_time": 2.6493751508198450e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8555103270286959e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 305, + "real_time": 2.2955576553330068e+06, + "cpu_time": 2.6486990918034487e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8549054234271877e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 305, + "real_time": 2.2952685964706005e+06, + "cpu_time": 2.6426742098360816e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8552649611803044e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 305, + "real_time": 2.3040660153158377e+06, + "cpu_time": 2.6573269278686461e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8443629463896431e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2987660645033987e+06, + "cpu_time": 2.6498254052459570e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8509305397313830e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time_median", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2955576553330072e+06, + "cpu_time": 2.6493751508198450e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.8549054234271877e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7508615318786742e+03, + "cpu_time": 5.2532707576844141e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8876140020809791e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0667007422980211e-03, + "cpu_time": 1.9824969400943629e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0651551905700640e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 295, + "real_time": 2.3816872561760880e+06, + "cpu_time": 2.7483247016951251e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5033254118528701e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 295, + "real_time": 2.3807082735633445e+06, + "cpu_time": 2.7478425796609093e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5055884610262185e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 295, + "real_time": 2.3899053207645980e+06, + "cpu_time": 2.7565973355934657e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4844013635680914e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 295, + "real_time": 2.3737257291262937e+06, + "cpu_time": 2.7403847016947079e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5217836834183939e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 295, + "real_time": 2.3871533477963028e+06, + "cpu_time": 2.7527360915254867e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.4907239252559513e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3826359854853256e+06, + "cpu_time": 2.7491770820339392e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5011645690243058e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time_median", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3816872561760875e+06, + "cpu_time": 2.7483247016951251e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 5.5033254118528701e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2701400818531611e+03, + "cpu_time": 6.0733045432227427e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4486200675340998e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6315979948468627e-03, + "cpu_time": 2.2091354474443295e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6332970943841972e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 282, + "real_time": 2.4935254490280403e+06, + "cpu_time": 2.8903059539007340e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0512986747425497e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 282, + "real_time": 2.4827192099897047e+06, + "cpu_time": 2.8834245390071501e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0558745384706113e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 282, + "real_time": 2.4830562373002372e+06, + "cpu_time": 2.8824103723404380e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0557312237318572e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 282, + "real_time": 2.4942142842667745e+06, + "cpu_time": 2.8908952907799087e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0510083341819310e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 282, + "real_time": 2.4884950851770579e+06, + "cpu_time": 2.8848890602838644e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0534238205310673e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4884020531523628e+06, + "cpu_time": 2.8863850432624193e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0534673183316034e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time_median", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4884950851770579e+06, + "cpu_time": 2.8848890602838648e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.0534238205310673e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4980070098135884e+03, + "cpu_time": 3.9533666776177392e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3276321971587351e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2094528506149524e-03, + "cpu_time": 1.3696601868298671e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2094963523358762e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 271, + "real_time": 2.5797819227282633e+06, + "cpu_time": 3.0497255793357319e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0322958129946744e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 271, + "real_time": 2.5816824754555728e+06, + "cpu_time": 3.0492316605170509e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0307997012974352e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 271, + "real_time": 2.5810070707019405e+06, + "cpu_time": 3.0582209520294559e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0313311263320664e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 271, + "real_time": 2.5813565761318286e+06, + "cpu_time": 3.0520235719564394e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0310560921639401e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 271, + "real_time": 2.5778324076926466e+06, + "cpu_time": 3.0472416273062085e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0338327597847104e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5803320905420510e+06, + "cpu_time": 3.0512886782289771e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0318630985145655e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time_median", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5810070707019409e+06, + "cpu_time": 3.0497255793357315e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0313311263320664e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5716080394562437e+03, + "cpu_time": 4.2323075862767109e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2381311146239722e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.0907200480775944e-04, + "cpu_time": 1.3870557762935818e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0935754752824295e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 256, + "real_time": 2.7440267513156869e+06, + "cpu_time": 3.3439942148438818e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8213038538973284e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 256, + "real_time": 2.7412499021011172e+06, + "cpu_time": 3.3463739023442683e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8251747832121617e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 256, + "real_time": 2.7387167519918876e+06, + "cpu_time": 3.3400703710935176e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8287128423827088e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 256, + "real_time": 2.7424661320765154e+06, + "cpu_time": 3.3503936757819373e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8234783931718016e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 256, + "real_time": 2.7429411275079474e+06, + "cpu_time": 3.3437685859372122e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8228162809774411e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7418801329986313e+06, + "cpu_time": 3.3449201500001634e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8242972307282877e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time_median", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7424661320765154e+06, + "cpu_time": 3.3439942148438818e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8234783931718016e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0298549730642633e+03, + "cpu_time": 3.8005490231820381e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8324784482851764e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.4031499358228031e-04, + "cpu_time": 1.1362151718873924e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.4065332200807198e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 229, + "real_time": 3.0653476501575075e+06, + "cpu_time": 3.9903235109174633e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8414817480563474e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 229, + "real_time": 3.0620756374029485e+06, + "cpu_time": 3.9918360960696968e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8487922844997609e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 229, + "real_time": 3.0609940607751431e+06, + "cpu_time": 3.9927178777291900e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8512122479222751e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 229, + "real_time": 3.0706520995540000e+06, + "cpu_time": 4.0031970917034294e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8296633158298945e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 229, + "real_time": 3.0651303490127670e+06, + "cpu_time": 3.9945388427945166e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8419667720671701e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0648399593804730e+06, + "cpu_time": 3.9945226838428592e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8426232736750901e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time_median", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0651303490127665e+06, + "cpu_time": 3.9927178777291900e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 6.8419667720671701e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7601682653784469e+03, + "cpu_time": 5.0831521977589709e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.3893050310499500e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2268726312673526e-03, + "cpu_time": 1.2725305624923413e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2260363745765820e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 195, + "real_time": 3.5799258830360114e+06, + "cpu_time": 5.0600514461539444e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1716175521608720e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 195, + "real_time": 3.5846488013003883e+06, + "cpu_time": 5.0655058205125108e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1700738991441643e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 195, + "real_time": 3.5839538280971539e+06, + "cpu_time": 5.0611389743596353e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1703007910196495e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 195, + "real_time": 3.6009614272281909e+06, + "cpu_time": 5.0844785076932544e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1647733764336736e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 195, + "real_time": 3.5862553542336593e+06, + "cpu_time": 5.0650287743594209e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1695497352268920e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5871490587790809e+06, + "cpu_time": 5.0672407046157531e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1692630707970505e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time_median", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5846488013003878e+06, + "cpu_time": 5.0650287743594209e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1700738991441643e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0655948999868633e+03, + "cpu_time": 9.9242236800029532e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6228524706662400e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2484694022534050e-03, + "cpu_time": 1.9585064650595677e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2431671162575265e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 145, + "real_time": 4.8348013793343101e+06, + "cpu_time": 7.5289033241382455e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7350470767746418e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 145, + "real_time": 4.8697323217217261e+06, + "cpu_time": 7.5744489103444414e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7226014585200346e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 145, + "real_time": 4.8674205960384728e+06, + "cpu_time": 7.5452717172413087e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7234195883600800e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 145, + "real_time": 4.8704881895461986e+06, + "cpu_time": 7.5740996482761996e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7223341220711589e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 145, + "real_time": 4.9360686302955812e+06, + "cpu_time": 7.6655197310346607e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.6994512492217262e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8757022233872581e+06, + "cpu_time": 7.5776486662069727e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7205706989895284e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time_median", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8697323217217270e+06, + "cpu_time": 7.5740996482761996e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.7226014585200346e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6906616458262200e+04, + "cpu_time": 5.2840640603351618e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2951317386187399e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5694976369213868e-03, + "cpu_time": 6.9732238760287172e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5273381057770891e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 88, + "real_time": 7.6893580103801051e+06, + "cpu_time": 1.3137896863637168e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1818747387430668e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 88, + "real_time": 7.7454474733465100e+06, + "cpu_time": 1.3270335977272846e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1660744660309744e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 88, + "real_time": 7.7161018429747364e+06, + "cpu_time": 1.3176711806819826e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1743124107770972e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 88, + "real_time": 7.7026490630073985e+06, + "cpu_time": 1.3159993056818433e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1781098765844016e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 88, + "real_time": 7.7213491736487905e+06, + "cpu_time": 1.3194362045453547e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1728347757224636e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7149811126715075e+06, + "cpu_time": 1.3187859950000364e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1746412535716009e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time_median", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7161018429747354e+06, + "cpu_time": 1.3176711806819823e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.1743124107770972e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1086904643336184e+04, + "cpu_time": 5.0600709359771805e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9396128218596876e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7332412530086296e-03, + "cpu_time": 3.8369158871580530e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7313069740144721e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.3753458017519876e+07, + "cpu_time": 2.4960114901962999e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4397087595902505e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.3761392173667749e+07, + "cpu_time": 2.4983173431373157e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4383021409858508e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.3771312402597830e+07, + "cpu_time": 2.4973215156861961e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4365456986997309e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3789373541287348e+07, + "cpu_time": 2.5012161549016930e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4333543434394069e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.3766292646965560e+07, + "cpu_time": 2.5018882568626091e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4374341633218331e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3768365756407673e+07, + "cpu_time": 2.4989509521568231e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4370690212074142e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time_median", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3766292646965558e+07, + "cpu_time": 2.4983173431373157e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.4374341633218331e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3464574509701915e+04, + "cpu_time": 2.5226901628189076e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3819545157706100e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.7793556242763407e-04, + "cpu_time": 1.0094996705083772e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.7738492223355317e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 28, + "real_time": 2.5148832837917976e+07, + "cpu_time": 4.8011087214279413e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6684683314136586e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 28, + "real_time": 2.5328645349613257e+07, + "cpu_time": 4.8147259321426973e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6495244050240803e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 28, + "real_time": 2.3902112751134805e+07, + "cpu_time": 4.5728522107140183e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.8076540638364224e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 28, + "real_time": 2.5153198479009528e+07, + "cpu_time": 4.7953267750001155e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6680051865373182e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 28, + "real_time": 2.5312073595289674e+07, + "cpu_time": 4.8101873178568766e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6512590423444524e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4968972602593049e+07, + "cpu_time": 4.7588401914283298e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6889822058311863e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time_median", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5153198479009524e+07, + "cpu_time": 4.8011087214279413e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.6680051865373182e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.0240385054057802e+05, + "cpu_time": 1.0424630903165148e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6939934591537811e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4126096821381345e-02, + "cpu_time": 2.1905822603461439e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4894153054034855e-02, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 15, + "real_time": 4.5620814214150108e+07, + "cpu_time": 8.7420578266664967e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9420283331631107e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 15, + "real_time": 4.5616771529118218e+07, + "cpu_time": 8.7524696199989185e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9422890638879557e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 15, + "real_time": 4.5621955146392189e+07, + "cpu_time": 8.7431549533327296e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9419547577327805e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 15, + "real_time": 4.5761742194493614e+07, + "cpu_time": 8.7623395866664097e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9329680550525470e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 15, + "real_time": 4.4686493277549744e+07, + "cpu_time": 8.5619994799996376e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.0035412975094709e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5461555272340775e+07, + "cpu_time": 8.7124042933328375e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9525563014691734e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time_median", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5620814214150108e+07, + "cpu_time": 8.7431549533327296e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.9420283331631107e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3761195724981243e+05, + "cpu_time": 8.4476782201634091e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8774190834404361e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.6259785796650801e-03, + "cpu_time": 9.6961503802434745e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.7455180854930708e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 8.6463508196175098e+07, + "cpu_time": 1.6541374574998713e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1046098128583083e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 8.6581669747829437e+07, + "cpu_time": 1.6549916350001583e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1003728246616492e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 8.5272543132305145e+07, + "cpu_time": 1.6218351662499231e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1479705675425591e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 8.6576090194284916e+07, + "cpu_time": 1.6547820162500671e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1005726338254075e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 8.6364588700234890e+07, + "cpu_time": 1.6524189225000897e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1081657429264169e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6251679994165897e+07, + "cpu_time": 1.6476330395000219e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1123383163628683e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time_median", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6463508196175098e+07, + "cpu_time": 1.6541374574998713e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.1046098128583083e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.5463135792756546e+05, + "cpu_time": 1.4456825292428460e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0176730017879508e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.4303832454635206e-03, + "cpu_time": 8.7742992194520548e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4828203000303576e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.6793927177786827e+08, + "cpu_time": 3.1333452475001878e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1968157674883466e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.6651555150747299e+08, + "cpu_time": 3.0795347775000435e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2241487785355949e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.6801107302308083e+08, + "cpu_time": 3.1339058750000960e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1954495756731844e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.6626326739788055e+08, + "cpu_time": 3.0783957150003970e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2290410287392426e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.6781470179557800e+08, + "cpu_time": 3.1314683299996203e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1991887853424463e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6730877310037616e+08, + "cpu_time": 3.1113299890000689e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.2089287871557627e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time_median", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6781470179557800e+08, + "cpu_time": 3.1314683299996203e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.1991887853424463e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.4690617510872183e+05, + "cpu_time": 2.9561363686196841e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6274461758404179e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.0619352435309793e-03, + "cpu_time": 9.5011984555509604e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0716182370718994e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 3.2104605436325073e+08, + "cpu_time": 5.6116946300005567e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3445102638922458e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 3.1965814530849457e+08, + "cpu_time": 5.5927035700005949e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3590316397655282e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 3.2057686150074005e+08, + "cpu_time": 5.6014630949994171e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3494052533093419e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 3.2020078599452972e+08, + "cpu_time": 5.5981792350007713e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3533391264640546e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 3.2004632055759430e+08, + "cpu_time": 5.6032971850004292e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3549575640466504e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2030563354492193e+08, + "cpu_time": 5.6014675430003536e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3522487694955640e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time_median", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2020078599452972e+08, + "cpu_time": 5.6014630949994171e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.3533391264640546e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2897656911325490e+05, + "cpu_time": 6.9921582932018279e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5340723844055384e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrGPU_fine/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6514744472612202e-03, + "cpu_time": 1.2482725713441462e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6508537298197856e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + } + ] +} diff --git a/results/implicit_managed_GPUWrHost_coarse.json b/results/implicit_managed_GPUWrHost_coarse.json new file mode 100644 index 0000000..971c807 --- /dev/null +++ b/results/implicit_managed_GPUWrHost_coarse.json @@ -0,0 +1,3194 @@ +{ + "context": { + "date": "2023-10-10T13:15:27-04:00", + "host_name": "frontier00741", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [2.33,5.29,7.95], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8648394535476754e+06, + "cpu_time": 2.0644529126984130e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1964357265220652e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8684047962581709e+06, + "cpu_time": 2.0679687645502645e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1922444259418538e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8602764727595502e+06, + "cpu_time": 2.0639859497354492e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.2018232558324831e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8651400083043396e+06, + "cpu_time": 2.0643565158730149e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1960817856906136e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8699616089059168e+06, + "cpu_time": 2.0694324391534382e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1904193008521181e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8657244679551304e+06, + "cpu_time": 2.0660393164021161e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1954008989678267e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8651400083043394e+06, + "cpu_time": 2.0644529126984130e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1960817856906136e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7397532834086023e+03, + "cpu_time": 2.4900231502644283e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4034300334591799e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0044510042297099e-03, + "cpu_time": 1.2052157625928701e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0057521318905644e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8656801196730207e+06, + "cpu_time": 2.0663804444444450e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3908920471510040e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8706439345637420e+06, + "cpu_time": 2.0711227777777787e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3792406714271242e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8667057357710309e+06, + "cpu_time": 2.0666942883597903e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3884795782321552e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8723260666500586e+06, + "cpu_time": 2.0722194100529091e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3753062812702386e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8707180428483449e+06, + "cpu_time": 2.0712396904761910e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3790671883010799e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8692147799012393e+06, + "cpu_time": 2.0695313222222228e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3825971532763205e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8706439345637423e+06, + "cpu_time": 2.0711227777777785e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3792406714271242e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8623204375554574e+03, + "cpu_time": 2.7682964512873241e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7136358537645701e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5312956372550664e-03, + "cpu_time": 1.3376441426915834e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5318852312824652e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8669796154811848e+06, + "cpu_time": 2.0687449312169291e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7756716057005692e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8716856675430709e+06, + "cpu_time": 2.0728230820105795e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7536065933052693e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8670477519177082e+06, + "cpu_time": 2.0684052989418025e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7753513444803096e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8711836417895500e+06, + "cpu_time": 2.0726103835978818e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7559551259922199e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8728928738537002e+06, + "cpu_time": 2.0746875291005308e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7479643009629101e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8699579101170427e+06, + "cpu_time": 2.0714542449735447e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7617097940882575e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8711836417895500e+06, + "cpu_time": 2.0726103835978813e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7559551259922199e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7586607109728689e+03, + "cpu_time": 2.7523149588312149e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2930010919671047e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4752528364663571e-03, + "cpu_time": 1.3286873053121023e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4757406058341770e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 377, + "real_time": 1.8698781381274108e+06, + "cpu_time": 2.0725376286472173e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7524136643907454e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 377, + "real_time": 1.8736527795480029e+06, + "cpu_time": 2.0770099708222745e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7488832700317562e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 377, + "real_time": 1.8674730054913114e+06, + "cpu_time": 2.0707286339522512e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7546706112294834e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 377, + "real_time": 1.8721270465160182e+06, + "cpu_time": 2.0752581432360753e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7503085627110850e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 377, + "real_time": 1.8668945206576879e+06, + "cpu_time": 2.0702803474801027e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7552143218276825e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8700050980680860e+06, + "cpu_time": 2.0731629448275845e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7522980860381510e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8698781381274108e+06, + "cpu_time": 2.0725376286472171e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7524136643907454e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9116811667282482e+03, + "cpu_time": 2.9075643847426691e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7279038077569745e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5570445073846717e-03, + "cpu_time": 1.4024775003802115e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5567578538675541e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 381, + "real_time": 1.8698409263460452e+06, + "cpu_time": 2.0779340682414682e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5048970784946583e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 381, + "real_time": 1.8733352912514131e+06, + "cpu_time": 2.0814824540682368e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.4983593330065906e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 381, + "real_time": 1.8688170419986974e+06, + "cpu_time": 2.0766469842519707e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5068173356290318e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 381, + "real_time": 1.8741512651672265e+06, + "cpu_time": 2.0824730787401628e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.4968362062361255e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 381, + "real_time": 1.8688288081579127e+06, + "cpu_time": 2.0769759475065635e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5067952566826187e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8709946665842594e+06, + "cpu_time": 2.0791025065616805e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5027410420098051e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8698409263460450e+06, + "cpu_time": 2.0779340682414682e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5048970784946583e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5596262687088097e+03, + "cpu_time": 2.6898883937498808e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7897750185290541e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3680564217650795e-03, + "cpu_time": 1.2937738208003457e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3674362338189790e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 380, + "real_time": 1.8707915593710679e+06, + "cpu_time": 2.0875385184210527e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0062321664560243e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 380, + "real_time": 1.8679532550250818e+06, + "cpu_time": 2.0883191947368418e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0168779463509679e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 380, + "real_time": 1.8768091714407904e+06, + "cpu_time": 2.0933850394736882e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9837680886532828e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 380, + "real_time": 1.8694807367864999e+06, + "cpu_time": 2.0896496894736821e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0111447216783389e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 380, + "real_time": 1.8713131650369368e+06, + "cpu_time": 2.0883172973684226e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0042792648986056e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8712695775320753e+06, + "cpu_time": 2.0894419478947376e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0044604376074448e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8707915593710677e+06, + "cpu_time": 2.0883191947368416e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0062321664560243e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3584017066353153e+03, + "cpu_time": 2.3312863928606175e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2552084775048327e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7947182741380027e-03, + "cpu_time": 1.1157459508313002e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7920130874971173e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 379, + "real_time": 1.8744517857544271e+06, + "cpu_time": 2.1072382189973542e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3985102310566637e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 379, + "real_time": 1.8772195086576103e+06, + "cpu_time": 2.1103279445910305e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3964483044790950e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 379, + "real_time": 1.8718094428823669e+06, + "cpu_time": 2.1076312796833832e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.4004844403195712e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 379, + "real_time": 1.8721901192569411e+06, + "cpu_time": 2.1151984538258621e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.4001996768578348e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 379, + "real_time": 1.8712994220588568e+06, + "cpu_time": 2.1202183957783626e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.4008661409812319e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8733940557220406e+06, + "cpu_time": 2.1121228585751988e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3993017587388796e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8721901192569416e+06, + "cpu_time": 2.1103279445910305e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.4001996768578348e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4546300143011899e+03, + "cpu_time": 5.5283973651502165e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8317342704388101e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3102582485536554e-03, + "cpu_time": 2.6174601267653424e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3090344945250838e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 370, + "real_time": 1.8956794228083834e+06, + "cpu_time": 2.1774812189189130e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7656996942198461e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 370, + "real_time": 1.8953576459819954e+06, + "cpu_time": 2.1773256567567657e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7661692299152511e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 370, + "real_time": 1.8947301112741509e+06, + "cpu_time": 2.1767712108107936e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7670853853028786e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 370, + "real_time": 1.8932576940199865e+06, + "cpu_time": 2.1770475648648702e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7692373925430632e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 370, + "real_time": 1.8912850780808995e+06, + "cpu_time": 2.1754135270270268e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7721257153469360e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8940619904330834e+06, + "cpu_time": 2.1768078356756740e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7680634834655952e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8947301112741511e+06, + "cpu_time": 2.1770475648648706e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7670853853028786e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8096499094196915e+03, + "cpu_time": 8.2530172276591088e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6463258726014831e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.5543330606930629e-04, + "cpu_time": 3.7913393605078601e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.5602065791074360e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 369, + "real_time": 1.8944554258047082e+06, + "cpu_time": 2.2579298970189658e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5349731944978130e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 369, + "real_time": 1.9002570242237370e+06, + "cpu_time": 2.2614912628726200e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5180745900852418e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 369, + "real_time": 1.8930837982409915e+06, + "cpu_time": 2.2574300704607060e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5389835408993089e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 369, + "real_time": 1.9000500786373124e+06, + "cpu_time": 2.2610773983739880e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5186755959191513e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 369, + "real_time": 1.8917754853111478e+06, + "cpu_time": 2.2559703414634299e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5428141877392840e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8959243624435794e+06, + "cpu_time": 2.2587797940379423e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5307042218281591e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8944554258047082e+06, + "cpu_time": 2.2579298970189658e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5349731944978130e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9759669736819496e+03, + "cpu_time": 2.4014657695113374e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1593295105361547e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0971126551470056e-03, + "cpu_time": 1.0631694934805135e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0961697896636779e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 363, + "real_time": 1.9410245251051630e+06, + "cpu_time": 2.4562165812672204e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0804356013411927e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 363, + "real_time": 1.9417293092361640e+06, + "cpu_time": 2.4568358374655596e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0800434386114180e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 363, + "real_time": 1.9368864660030717e+06, + "cpu_time": 2.4596216997245299e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0827438968726182e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 363, + "real_time": 1.9406303888463902e+06, + "cpu_time": 2.4560532176308557e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0806550345976260e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 363, + "real_time": 1.9471585691570414e+06, + "cpu_time": 2.4562018429752034e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0770319547770028e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9414858516695662e+06, + "cpu_time": 2.4569858358126739e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0801819852399716e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9410245251051630e+06, + "cpu_time": 2.4562165812672204e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0804356013411927e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6857150807125340e+03, + "cpu_time": 1.5038660231623762e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0489860370046450e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8983991449348090e-03, + "cpu_time": 6.1207761202455467e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8968896584119993e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 344, + "real_time": 2.0414221974977746e+06, + "cpu_time": 2.8616373808139651e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0545989972780101e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 344, + "real_time": 2.0438385044419488e+06, + "cpu_time": 2.8610481337209446e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0521699688524146e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 344, + "real_time": 2.0457342839637382e+06, + "cpu_time": 2.8615175319767543e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0502682253891125e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 344, + "real_time": 2.0424601358637691e+06, + "cpu_time": 2.8599081453488404e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0535548901796327e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 344, + "real_time": 2.0435704050320310e+06, + "cpu_time": 2.8610133023255845e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0524391964534535e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0434051053598523e+06, + "cpu_time": 2.8610248988372181e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0526062556305246e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0435704050320308e+06, + "cpu_time": 2.8610481337209442e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0524391964534535e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6174920907465385e+03, + "cpu_time": 6.8294111492444199e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6244666133595975e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.9156702041306248e-04, + "cpu_time": 2.3870505817757961e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.9141657534342348e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 335, + "real_time": 2.0836719402125967e+06, + "cpu_time": 3.5708082447761279e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.0258775088866014e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 335, + "real_time": 2.0880754918916456e+06, + "cpu_time": 3.5728907194029759e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.0173873179271536e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 335, + "real_time": 2.0808434023507941e+06, + "cpu_time": 3.5710297074626936e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.0313499759391441e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 335, + "real_time": 2.0901964966958358e+06, + "cpu_time": 3.5742951731343134e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.0133107166051788e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 335, + "real_time": 2.0851499679039663e+06, + "cpu_time": 3.5689929283582182e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.0230238252035146e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0855874598109678e+06, + "cpu_time": 3.5716033546268656e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.0221898689123187e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0851499679039665e+06, + "cpu_time": 3.5710297074626936e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 4.0230238252035146e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6674316737625495e+03, + "cpu_time": 2.0419837773965999e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.0729338690192774e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7584645786539953e-03, + "cpu_time": 5.7172747773105693e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7584783661473299e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 199, + "real_time": 3.5050070678450204e+06, + "cpu_time": 6.3639322864321750e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.7866425588451433e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 199, + "real_time": 3.5085069617003803e+06, + "cpu_time": 6.3734579949748823e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.7818676671141644e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 199, + "real_time": 3.5099926657375679e+06, + "cpu_time": 6.3735384472361691e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.7798436058767490e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 199, + "real_time": 3.5046290175699107e+06, + "cpu_time": 6.3697181658291444e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.7871589020949278e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 199, + "real_time": 3.5031165929988553e+06, + "cpu_time": 6.3590011859296169e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.7892256950653782e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5062504611703469e+06, + "cpu_time": 6.3679296160803977e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.7849476857992725e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5050070678450214e+06, + "cpu_time": 6.3697181658291453e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 4.7866425588451433e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8762877849359220e+03, + "cpu_time": 6.3426410033092479e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9243337774455426e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.2033152417065186e-04, + "cpu_time": 9.9602875435254629e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2014141744791644e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 101, + "real_time": 7.0638617288579447e+06, + "cpu_time": 1.2958763663366394e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.7501541349429750e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 101, + "real_time": 7.0678948996459497e+06, + "cpu_time": 1.2954681762376240e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.7474435424444180e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 101, + "real_time": 7.0616721940955315e+06, + "cpu_time": 1.2955119108910920e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.7516269628114195e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 101, + "real_time": 7.0767075856133262e+06, + "cpu_time": 1.2964943287128694e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.7415315094006243e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 101, + "real_time": 7.0726300018037306e+06, + "cpu_time": 1.2952758544554427e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.7442651448531342e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0685532820032965e+06, + "cpu_time": 1.2957253273267336e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.7470042588905144e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0678948996459497e+06, + "cpu_time": 1.2955119108910920e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.7474435424444180e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1827546592204653e+03, + "cpu_time": 4.8163135105743240e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1514399135143459e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.7468459422409726e-04, + "cpu_time": 3.7170790822705203e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7453890645647629e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50, + "real_time": 1.4118497483432293e+07, + "cpu_time": 2.5710182859999977e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7532582046177778e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50, + "real_time": 1.3772371094673872e+07, + "cpu_time": 2.4737117079999961e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.8727167993572807e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50, + "real_time": 1.4122546054422855e+07, + "cpu_time": 2.5712053560000073e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7518955676538973e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50, + "real_time": 1.4108270965516567e+07, + "cpu_time": 2.5685967039999865e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7567036502224455e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50, + "real_time": 1.4127106349915266e+07, + "cpu_time": 2.5707315079999942e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7503616337115288e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4049758389592171e+07, + "cpu_time": 2.5510527123999964e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7769871711125860e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4118497483432293e+07, + "cpu_time": 2.5707315079999942e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7532582046177778e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5522020411590324e+05, + "cpu_time": 4.3247632225270953e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3565741385072604e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1047891345297994e-02, + "cpu_time": 1.6952857153854792e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1213289771635885e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 26, + "real_time": 2.7222325308964804e+07, + "cpu_time": 4.8701150538461275e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.9304284801783514e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 26, + "real_time": 2.7165949774476197e+07, + "cpu_time": 4.8674630461538672e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.9406602424813595e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 26, + "real_time": 2.7202559563402947e+07, + "cpu_time": 4.8684332076922812e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.9340109958097572e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 26, + "real_time": 2.7197267287052594e+07, + "cpu_time": 4.8694790576923065e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.9349710977725716e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 26, + "real_time": 2.6773948652239945e+07, + "cpu_time": 4.7087592769231014e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0129971392460690e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7112410117227297e+07, + "cpu_time": 4.8368499284615375e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.9506135910976210e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7197267287052594e+07, + "cpu_time": 4.8684332076922819e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.9349710977725716e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9028378114737102e+05, + "cpu_time": 7.1611992980510555e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5066182779314891e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.0183277814340889e-03, + "cpu_time": 1.4805502349602206e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.0831993113686384e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10, + "real_time": 5.2709759026765823e+07, + "cpu_time": 9.0856502500000149e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0927088447452297e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10, + "real_time": 5.2530158311128616e+07, + "cpu_time": 9.0797320400000110e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.1101208264040470e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10, + "real_time": 4.9138173088431358e+07, + "cpu_time": 8.7876537600000411e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.4628700891445637e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10, + "real_time": 5.2662638574838638e+07, + "cpu_time": 9.0868765400000483e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0972655997577410e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10, + "real_time": 5.2701373770833015e+07, + "cpu_time": 9.1383154000000387e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0935191398855448e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1948420554399490e+07, + "cpu_time": 9.0356455980000302e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.1712968999874258e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2662638574838638e+07, + "cpu_time": 9.0856502500000149e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.0972655997577410e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5726245446663951e+06, + "cpu_time": 1.4063225889301652e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6314374981992292e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.0272807678908543e-02, + "cpu_time": 1.5564162778157697e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1547937195468244e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 1.0337692871689796e+08, + "cpu_time": 1.7664541862500015e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1933339349850817e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 9.2233662493526936e+07, + "cpu_time": 1.7136764125000072e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.8207697437763376e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 1.0341954883188009e+08, + "cpu_time": 1.7669645537500146e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1911937159264059e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 9.2157389037311077e+07, + "cpu_time": 1.7140635262500048e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.8255872655272503e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 1.0356452316045761e+08, + "cpu_time": 1.7681576812499955e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1839268469203444e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8950410448014736e+07, + "cpu_time": 1.7458632720000046e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4429623014270840e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0337692871689796e+08, + "cpu_time": 1.7664541862500018e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.1933339349850817e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1667888923737342e+06, + "cpu_time": 2.9212618331746161e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4711000969295853e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.2322014274145535e-02, + "cpu_time": 1.6732477737664488e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.3772260484323065e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.8650595098733902e+08, + "cpu_time": 3.0232801874999994e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7571451115406561e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.8674126639962196e+08, + "cpu_time": 3.0159002174999827e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7498904484358454e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.8665195256471634e+08, + "cpu_time": 3.0156870475000018e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7526417979887466e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.8721435591578484e+08, + "cpu_time": 3.0193090650000441e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7353605109375496e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.8669479712843895e+08, + "cpu_time": 3.0144142099999982e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7513216250011845e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8676166459918022e+08, + "cpu_time": 3.0177181455000055e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7492718987807970e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8669479712843895e+08, + "cpu_time": 3.0159002174999827e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.7513216250011845e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6797525777389959e+05, + "cpu_time": 3.5990813807237387e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2379365251645390e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_coarse/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4348515170338433e-03, + "cpu_time": 1.1926499451549698e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4328660515971585e-03, + "dst_gpu": NaN, + "src_numa": NaN + } + ] +} diff --git a/results/implicit_managed_GPUWrHost_fine.json b/results/implicit_managed_GPUWrHost_fine.json new file mode 100644 index 0000000..bb7e159 --- /dev/null +++ b/results/implicit_managed_GPUWrHost_fine.json @@ -0,0 +1,3194 @@ +{ + "context": { + "date": "2023-10-10T13:13:12-04:00", + "host_name": "frontier00741", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [1.55,6.82,8.79], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 381, + "real_time": 1.8642109173392509e+06, + "cpu_time": 2.0637587427821513e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1971762754431968e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 381, + "real_time": 1.8685784105835126e+06, + "cpu_time": 2.0681197559055125e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1920407389920107e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 381, + "real_time": 1.8645023822708342e+06, + "cpu_time": 2.0638868293963235e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1968328058725013e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 381, + "real_time": 1.8690428719489183e+06, + "cpu_time": 2.0680988766404206e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1914960119287972e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 381, + "real_time": 1.8645859721576273e+06, + "cpu_time": 2.0638559842519695e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1967343212714759e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8661841108600288e+06, + "cpu_time": 2.0655440377952754e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1948560307015963e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8645859721576273e+06, + "cpu_time": 2.0638868293963238e+06, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.1967343212714759e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4073292857951005e+03, + "cpu_time": 2.3422568196641673e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8299622243779600e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2899741626702019e-03, + "cpu_time": 1.1339660529166207e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2893612085679027e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 382, + "real_time": 1.8640420339912900e+06, + "cpu_time": 2.0642850497382216e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3947506819142243e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 382, + "real_time": 1.8679696125013717e+06, + "cpu_time": 2.0686371884816759e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3855103130024746e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 382, + "real_time": 1.8640336759540213e+06, + "cpu_time": 2.0646105549738223e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3947703872931888e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 382, + "real_time": 1.8696827047417019e+06, + "cpu_time": 2.0702544659685839e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3814920998222157e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 382, + "real_time": 1.8637002851572973e+06, + "cpu_time": 2.0643254371727745e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3955565523286872e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8658856624691368e+06, + "cpu_time": 2.0664225392670159e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3904160068721585e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8640420339912903e+06, + "cpu_time": 2.0646105549738221e+06, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.3947506819142243e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7552255700931223e+03, + "cpu_time": 2.8212727246124859e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4783715842632137e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4766315136625883e-03, + "cpu_time": 1.3652932403714609e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4755712383798835e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8697716641038405e+06, + "cpu_time": 2.0713753386243409e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7625672773539741e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8706090902529592e+06, + "cpu_time": 2.0723318201058202e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7586444893114567e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8649608961904913e+06, + "cpu_time": 2.0672095925925935e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7851707955202628e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8690812225580134e+06, + "cpu_time": 2.0705769656084643e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7658041834998243e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8656865761614719e+06, + "cpu_time": 2.0677561984126994e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7817537036199328e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8680218898533557e+06, + "cpu_time": 2.0698499830687835e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7707880898610912e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8690812225580134e+06, + "cpu_time": 2.0705769656084641e+06, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.7658041834998243e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5348014644936202e+03, + "cpu_time": 2.2566747836321892e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1905828638377350e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3569441976360400e-03, + "cpu_time": 1.0902600681651415e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3574411462682950e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 378, + "real_time": 1.8717497887866567e+06, + "cpu_time": 2.0748227592592605e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7506613435357474e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 378, + "real_time": 1.8677782096905282e+06, + "cpu_time": 2.0704962566137551e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7543838893713899e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 378, + "real_time": 1.8633198050477805e+06, + "cpu_time": 2.0709903862433846e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7585816407484461e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 378, + "real_time": 1.8721258447118409e+06, + "cpu_time": 2.0751412698412663e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7503096863151141e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 378, + "real_time": 1.8672826856980841e+06, + "cpu_time": 2.0703938439153465e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7548494532176137e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8684512667869781e+06, + "cpu_time": 2.0723689031746027e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7537572026376624e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8677782096905280e+06, + "cpu_time": 2.0709903862433848e+06, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.7543838893713899e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6239914977442372e+03, + "cpu_time": 2.3987183956709405e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4033315679960935e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9395697186023575e-03, + "cpu_time": 1.1574765438703563e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9405944921437588e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 377, + "real_time": 1.8743483427226783e+06, + "cpu_time": 2.0823899257294498e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.4964685328876704e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 377, + "real_time": 1.8686924589981898e+06, + "cpu_time": 2.0770164244031827e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5070511300256439e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 377, + "real_time": 1.8642748509109421e+06, + "cpu_time": 2.0767835517241312e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5153614805229545e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 377, + "real_time": 1.8750968068955154e+06, + "cpu_time": 2.0828287586206871e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.4950728815171950e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 377, + "real_time": 1.8688389861311477e+06, + "cpu_time": 2.0767795623342199e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5067761581575304e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8702502891316947e+06, + "cpu_time": 2.0791596445623343e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5041460366221987e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8688389861311477e+06, + "cpu_time": 2.0770164244031827e+06, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.5067761581575304e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4835045020734542e+03, + "cpu_time": 3.1544062686980260e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4025218837114604e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3972751284294792e-03, + "cpu_time": 1.5171544315742201e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3978800529132691e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 377, + "real_time": 1.8668604523853823e+06, + "cpu_time": 2.0881156790450865e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0209854106943384e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 377, + "real_time": 1.8753986975900084e+06, + "cpu_time": 2.0920567851458844e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9890205303242892e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 377, + "real_time": 1.8701233660378580e+06, + "cpu_time": 2.0868578514588906e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0087354866698474e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 377, + "real_time": 1.8746777011028465e+06, + "cpu_time": 2.0916486419098147e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.9917084906323999e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 377, + "real_time": 1.8667982082470600e+06, + "cpu_time": 2.0875026180371314e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0212195094764829e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8707716850726311e+06, + "cpu_time": 2.0892363151193620e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0063338855594724e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8701233660378580e+06, + "cpu_time": 2.0881156790450867e+06, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.0087354866698474e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1283405518645059e+03, + "cpu_time": 2.4337723742178046e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5456965438759318e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2067580906882421e-03, + "cpu_time": 1.1649100470851996e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2061417128032062e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 376, + "real_time": 1.8788010299257177e+06, + "cpu_time": 2.1115044867021316e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3952728140156725e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 376, + "real_time": 1.8774150951859602e+06, + "cpu_time": 2.1100629547872362e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3963028244109988e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 376, + "real_time": 1.8778432199627440e+06, + "cpu_time": 2.1102690319148917e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3959844848240361e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 376, + "real_time": 1.8768721924487257e+06, + "cpu_time": 2.1096920744680809e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3967067179890648e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 376, + "real_time": 1.8769007000376773e+06, + "cpu_time": 2.1095148537234101e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3966855038987288e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8775664475121652e+06, + "cpu_time": 2.1102086803191500e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3961904690277001e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8774150951859602e+06, + "cpu_time": 2.1100629547872366e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3963028244109988e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.9826593924035581e+02, + "cpu_time": 7.8298617139174507e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9344680890539799e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.2515988730949220e-04, + "cpu_time": 3.7104679679041282e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2504717090546487e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 374, + "real_time": 1.8923375010054638e+06, + "cpu_time": 2.1558289759358303e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7705839984750491e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 374, + "real_time": 1.8900059120615020e+06, + "cpu_time": 2.1552819919786076e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7740019047249383e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 374, + "real_time": 1.8871566881475481e+06, + "cpu_time": 2.1526532914438434e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7781900850779188e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 374, + "real_time": 1.8873432254914276e+06, + "cpu_time": 2.1525315401069499e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7779155000462919e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 374, + "real_time": 1.8884131575487603e+06, + "cpu_time": 2.1538580106952023e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7763415961396277e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8890512968509402e+06, + "cpu_time": 2.1540307620320865e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7754066168927652e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8884131575487603e+06, + "cpu_time": 2.1538580106952027e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7763415961396277e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1578171623635790e+03, + "cpu_time": 1.4978387968138838e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1680925131693995e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1422755782019647e-03, + "cpu_time": 6.9536555522579479e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1414876990947977e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 369, + "real_time": 1.8962890803050452e+06, + "cpu_time": 2.2250173035230292e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5296210419105589e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 369, + "real_time": 1.9015962373765646e+06, + "cpu_time": 2.2277543035230306e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5141884454221034e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 369, + "real_time": 1.8969655858301041e+06, + "cpu_time": 2.2244229918699092e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5276490403021610e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 369, + "real_time": 1.9028291039498204e+06, + "cpu_time": 2.2286213224932295e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5106157343473768e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 369, + "real_time": 1.8925927939060016e+06, + "cpu_time": 2.2253821598916068e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5404205456996953e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8980545602735076e+06, + "cpu_time": 2.2262396162601612e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5244989615363801e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8969655858301041e+06, + "cpu_time": 2.2253821598916072e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.5276490403021610e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1675358645728356e+03, + "cpu_time": 1.8368591476200631e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2131635148816502e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1956881281496453e-03, + "cpu_time": 8.2509498717203917e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1959702107433570e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 360, + "real_time": 1.9543261011777858e+06, + "cpu_time": 2.4097100749999909e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0730819174630780e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 360, + "real_time": 1.9455685679632653e+06, + "cpu_time": 2.4033338666666555e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0779121509942057e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 360, + "real_time": 1.9471640342898252e+06, + "cpu_time": 2.4042864888888872e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0770289318562105e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 360, + "real_time": 1.9460758826527228e+06, + "cpu_time": 2.4028399694444346e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0776311544138472e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 360, + "real_time": 1.9453463920702536e+06, + "cpu_time": 2.4005272472222294e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0780352581671553e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9476961956307709e+06, + "cpu_time": 2.4041395294444398e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0767378825788994e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9460758826527228e+06, + "cpu_time": 2.4033338666666555e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.0776311544138472e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7720231197459862e+03, + "cpu_time": 3.4073122260307960e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0803133723167768e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9366588732922993e-03, + "cpu_time": 1.4172689165084251e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9320518075710399e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 340, + "real_time": 2.0704094878644410e+06, + "cpu_time": 2.7722752852941146e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0258330656735380e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 340, + "real_time": 2.0716474959900712e+06, + "cpu_time": 2.7731897617647154e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0246224360653014e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 340, + "real_time": 2.0744604561888778e+06, + "cpu_time": 2.7766769235293972e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0218770560252669e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 340, + "real_time": 2.0712452906403034e+06, + "cpu_time": 2.7746813264705851e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0250155879429307e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 340, + "real_time": 2.0695340585114216e+06, + "cpu_time": 2.7719678264705795e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0266900091593018e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0714593578390232e+06, + "cpu_time": 2.7737582247058782e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0248076309732678e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0712452906403032e+06, + "cpu_time": 2.7731897617647154e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.0250155879429307e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8639217053723871e+03, + "cpu_time": 1.9428526029288410e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8207606062522333e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.9981089820504985e-04, + "cpu_time": 7.0044050185190723e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9922646398613438e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 320, + "real_time": 2.1811461865581805e+06, + "cpu_time": 3.4093874343750002e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8459632149815283e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 320, + "real_time": 2.1872430563234957e+06, + "cpu_time": 3.4123032437500143e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8352427160519996e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 320, + "real_time": 2.1817523313075071e+06, + "cpu_time": 3.4103196187500013e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8448947112951050e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 320, + "real_time": 2.1864943031687289e+06, + "cpu_time": 3.4110451343750013e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8365560741882534e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 320, + "real_time": 2.1817643373651663e+06, + "cpu_time": 3.4101986749999868e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8448735531769681e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1836800429446162e+06, + "cpu_time": 3.4106508212500014e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8415060539387712e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1817643373651663e+06, + "cpu_time": 3.4103196187500013e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8448735531769681e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9334942114278406e+03, + "cpu_time": 1.0950183430871698e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1580285806730464e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3433718098518346e-03, + "cpu_time": 3.2105847255446880e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3427099966130259e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 216, + "real_time": 3.2361089014793159e+06, + "cpu_time": 5.5134452824074095e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.1843792995750742e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 216, + "real_time": 3.2371890890687979e+06, + "cpu_time": 5.5162472962963069e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.1826493721520891e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 216, + "real_time": 3.2355661301321727e+06, + "cpu_time": 5.5127945277777696e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.1852489874205265e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 216, + "real_time": 3.2370891810084381e+06, + "cpu_time": 5.5181132962962901e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.1828093271046228e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 216, + "real_time": 3.2362826912301695e+06, + "cpu_time": 5.5193440787037034e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.1841008962114735e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2364471985837789e+06, + "cpu_time": 5.5159888962962953e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.1838375764927568e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2362826912301695e+06, + "cpu_time": 5.5162472962963069e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.1841008962114735e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8562601524806314e+02, + "cpu_time": 2.8509233929824641e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0981812664947440e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1184526524890712e-04, + "cpu_time": 5.1684719577603088e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1184715961678404e-04, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 109, + "real_time": 6.4732682540876055e+06, + "cpu_time": 1.0804238862385331e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1835380032044468e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 109, + "real_time": 6.4751620301891360e+06, + "cpu_time": 1.0809530247706346e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1820219854822531e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 109, + "real_time": 6.4740700202934239e+06, + "cpu_time": 1.0807462055045953e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1828960599470339e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 109, + "real_time": 6.4651600358614679e+06, + "cpu_time": 1.0796706844036760e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1900388874950638e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 109, + "real_time": 6.4541333816398717e+06, + "cpu_time": 1.0782171330275230e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1989058818419495e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4683587444143016e+06, + "cpu_time": 1.0800021867889924e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1874801635941496e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4732682540876064e+06, + "cpu_time": 1.0804238862385331e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.1835380032044468e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8825578698447607e+03, + "cpu_time": 1.1103825810382636e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1316275942704696e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3732321011903092e-03, + "cpu_time": 1.0281299377176232e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3747768414268627e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 53, + "real_time": 1.3220714001020171e+07, + "cpu_time": 2.1857016037735783e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0760393118572550e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 53, + "real_time": 1.2844021027942874e+07, + "cpu_time": 2.0934149735848963e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.2249107856489010e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 53, + "real_time": 1.3203476801655203e+07, + "cpu_time": 2.1853233849056769e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0826661043996506e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 53, + "real_time": 1.3212889942498703e+07, + "cpu_time": 2.1853147471697967e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0790451061086321e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 53, + "real_time": 1.3221433702504860e+07, + "cpu_time": 2.1859389547169857e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0757630004441900e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3140507095124362e+07, + "cpu_time": 2.1671387328301869e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.1076848616917257e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3212889942498703e+07, + "cpu_time": 2.1853233849056773e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 5.0790451061086321e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6589960224703461e+05, + "cpu_time": 4.1213681788604805e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5590656944688700e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2625053283414745e-02, + "cpu_time": 1.9017555804921437e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2841563001787133e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 28, + "real_time": 2.5614134900804076e+07, + "cpu_time": 4.1924226035714440e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.2399867697966499e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 28, + "real_time": 2.5637142027595215e+07, + "cpu_time": 4.1937790392857209e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.2352843330013618e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 28, + "real_time": 2.5607856589236431e+07, + "cpu_time": 4.1915580178571336e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.2412714641808329e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 28, + "real_time": 2.5590417241411548e+07, + "cpu_time": 4.1881220392857037e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.2448432838681078e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 28, + "real_time": 2.5390291187380042e+07, + "cpu_time": 4.0955217571428597e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.2861830929576502e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5567968389285468e+07, + "cpu_time": 4.1722806914285727e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.2495137887609205e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5607856589236431e+07, + "cpu_time": 4.1915580178571329e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.2412714641808329e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0072306905112433e+05, + "cpu_time": 4.2960368929362652e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0782377765050851e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.9394240292213991e-03, + "cpu_time": 1.0296615234353561e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9589147874123905e-03, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 14, + "real_time": 5.0144832315189496e+07, + "cpu_time": 8.0899167857143283e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3532027849395666e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 14, + "real_time": 4.6333931918655120e+07, + "cpu_time": 7.9248396785714149e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.7934961460052910e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 14, + "real_time": 5.0158627065164700e+07, + "cpu_time": 8.0861248857142910e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3517305338373013e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 14, + "real_time": 5.0128799464021407e+07, + "cpu_time": 8.0903294142856911e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3549149165772915e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 14, + "real_time": 5.0147611647844315e+07, + "cpu_time": 8.0863024357142881e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3529060942135448e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9382760482175015e+07, + "cpu_time": 8.0555026400000021e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.4412500951146002e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0144832315189503e+07, + "cpu_time": 8.0863024357142881e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3532027849395666e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7043803500881467e+06, + "cpu_time": 7.3069139296834846e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9691482079611173e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.4513671035124739e-02, + "cpu_time": 9.0707113587184945e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6189261172338089e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 9.8970509134232998e+07, + "cpu_time": 1.5685448937500012e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4245544121819744e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 9.5839703455567360e+07, + "cpu_time": 1.5285414662499973e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.6017589020285416e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 9.8968537524342537e+07, + "cpu_time": 1.5682321325000003e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4246624778905106e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 9.5778675749897957e+07, + "cpu_time": 1.5280485050000080e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.6053281985428991e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 9.9263352341949940e+07, + "cpu_time": 1.5694194212500000e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4085510849013672e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7764155641198173e+07, + "cpu_time": 1.5525572837500015e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4929710151090584e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8968537524342537e+07, + "cpu_time": 1.5682321325000006e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4246624778905106e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7887888242944565e+06, + "cpu_time": 2.2153306547368839e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0115897567554784e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8296980243551086e-02, + "cpu_time": 1.4268914119458697e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8416076727384555e-02, + "dst_gpu": NaN, + "src_numa": NaN + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.9029619172215462e+08, + "cpu_time": 2.8913222500000304e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.6424766795529785e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.9031654298305511e+08, + "cpu_time": 2.8728711950000018e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.6418733083839216e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.9041225314140320e+08, + "cpu_time": 2.8706815399999994e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.6390374373786860e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.9010109826922417e+08, + "cpu_time": 2.8642057174999636e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.6482673365692501e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.9038891047239304e+08, + "cpu_time": 2.8705967300000167e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.6397288126489687e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9030299931764606e+08, + "cpu_time": 2.8739354865000027e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.6422767149067612e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9031654298305511e+08, + "cpu_time": 2.8706815399999994e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 5.6418733083839216e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2279142075804199e+05, + "cpu_time": 1.0244951549178448e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6426293870225116e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_numa": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_GPUWrHost_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.4524164725897738e-04, + "cpu_time": 3.5647813241817664e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4559566484904420e-04, + "dst_gpu": NaN, + "src_numa": NaN + } + ] +} diff --git a/results/implicit_managed_HostWrGPU_coarse.json b/results/implicit_managed_HostWrGPU_coarse.json new file mode 100644 index 0000000..43097d6 --- /dev/null +++ b/results/implicit_managed_HostWrGPU_coarse.json @@ -0,0 +1,2362 @@ +{ + "context": { + "date": "2023-10-09T20:21:17-04:00", + "host_name": "frontier07630", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.45,0.25,0.2], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 214287, + "real_time": 3.2599878340731461e+03, + "cpu_time": 2.8247938405969570e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.2564464066978772e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 214287, + "real_time": 3.2667609467680004e+03, + "cpu_time": 2.8394147357515863e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.2538413635844443e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 214287, + "real_time": 3.2883553085350713e+03, + "cpu_time": 2.8416749597969081e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.2456074893636496e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 214287, + "real_time": 3.2898393089643687e+03, + "cpu_time": 2.8444475261215103e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.2450456132732537e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 214287, + "real_time": 3.2805403967578141e+03, + "cpu_time": 2.8384025001983333e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.2485747787310016e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2770967590196806e+03, + "cpu_time": 2.8377467124930589e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.2499031303300455e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2805403967578141e+03, + "cpu_time": 2.8394147357515863e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.2485747787310016e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3233768216744014e+01, + "cpu_time": 7.6046249024031943e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0531343745600116e+06, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.0382598348126884e-03, + "cpu_time": 2.6798110165802175e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0428208010213533e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 207615, + "real_time": 3.3402384365293092e+03, + "cpu_time": 3.0199442146280385e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.4525195298668370e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 207615, + "real_time": 3.3385067890086093e+03, + "cpu_time": 3.0109528115020582e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.4537916253369870e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 207615, + "real_time": 3.3219550032513880e+03, + "cpu_time": 3.0248943866772630e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.4660177491814365e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 207615, + "real_time": 3.3141794186355614e+03, + "cpu_time": 3.0152991398020391e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.4718034135196652e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 207615, + "real_time": 3.3290678033862387e+03, + "cpu_time": 3.0392262740649754e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.4607489194624743e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3287894901622217e+03, + "cpu_time": 3.0220633653348748e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.4609762474734802e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3290678033862387e+03, + "cpu_time": 3.0199442146280379e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.4607489194624743e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1019824515062533e+01, + "cpu_time": 1.0911806348548444e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1521593766559791e+06, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.3104600178623803e-03, + "cpu_time": 3.6107139491892508e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3125713362838248e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 196248, + "real_time": 3.5325397711058486e+03, + "cpu_time": 3.3933548222147481e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6380227999162912e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 196248, + "real_time": 3.5211939688555703e+03, + "cpu_time": 3.3889647967367811e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6529671880941553e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 196248, + "real_time": 3.5488896905954425e+03, + "cpu_time": 3.3912279461701523e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6166551875133228e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 196248, + "real_time": 3.5495088510454761e+03, + "cpu_time": 3.3915751019628195e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6158498788287954e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 196248, + "real_time": 3.5190440514043025e+03, + "cpu_time": 3.3904940321430028e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6558098621873846e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5342352666013276e+03, + "cpu_time": 3.3911233398455003e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6358609833079901e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5325397711058481e+03, + "cpu_time": 3.3912279461701517e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6380227999162912e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4592444067012835e+01, + "cpu_time": 1.6007718133622737e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9134787406396758e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.1288830443496636e-03, + "cpu_time": 4.7204765292765927e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1275584999839309e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 167373, + "real_time": 4.1891683186650689e+03, + "cpu_time": 4.1499570926613011e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.8220776792377567e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 167373, + "real_time": 4.1622891924022242e+03, + "cpu_time": 4.1190468281622475e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.8725908953693523e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 167373, + "real_time": 4.1677331469229393e+03, + "cpu_time": 4.1155755253834260e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.8623076010019503e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 167373, + "real_time": 4.1631011513203275e+03, + "cpu_time": 4.0751517364807968e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.8710554485584927e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 167373, + "real_time": 4.1719457618608540e+03, + "cpu_time": 4.1608637497087399e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.8543686496499815e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1708475142342832e+03, + "cpu_time": 4.1241189864793024e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.8564800547635059e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1677331469229393e+03, + "cpu_time": 4.1190468281622475e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.8623076010019503e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0952263665109466e+01, + "cpu_time": 3.3594249678168899e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0578769750999209e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6259084341327616e-03, + "cpu_time": 8.1458003002109793e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6193371086740023e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 133552, + "real_time": 5.2446455912302699e+03, + "cpu_time": 5.6491762253653980e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2495791919588373e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 133552, + "real_time": 5.2299371705401654e+03, + "cpu_time": 5.6478708622109715e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2530934476452080e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 133552, + "real_time": 5.2561436294474624e+03, + "cpu_time": 5.5649825077123626e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2468456842167629e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 133552, + "real_time": 5.2106039969448384e+03, + "cpu_time": 5.4358405993171118e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2577428650963701e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 133552, + "real_time": 5.1688653258654604e+03, + "cpu_time": 5.6578995758955309e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2678991590678915e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2220391428056400e+03, + "cpu_time": 5.5911539541002759e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2550320695970140e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2299371705401654e+03, + "cpu_time": 5.6478708622109727e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2530934476452080e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4259172970145237e+01, + "cpu_time": 9.4664608536226133e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2676997843155384e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.5604971608349228e-03, + "cpu_time": 1.6931139674092467e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5876402560535886e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 96624, + "real_time": 7.3096157993870975e+03, + "cpu_time": 8.5993048361690855e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.7931448601031837e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 96624, + "real_time": 7.2756168860736734e+03, + "cpu_time": 8.5940946299056313e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.8015242150928280e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 96624, + "real_time": 7.2389803568470597e+03, + "cpu_time": 8.5822495771236916e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.8106417414991917e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 96624, + "real_time": 7.2713154495775452e+03, + "cpu_time": 8.5965814281130873e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.8025899290012943e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 96624, + "real_time": 7.2934735986917349e+03, + "cpu_time": 8.6039932091405673e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.7971135183585361e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2778004181154220e+03, + "cpu_time": 8.5952447360904119e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.8010028528110065e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2756168860736743e+03, + "cpu_time": 8.5965814281130885e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.8015242150928280e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6512670842663475e+01, + "cpu_time": 8.1394692645843725e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5683669296735242e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6429510730563428e-03, + "cpu_time": 9.4697353181901920e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6470608136026060e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 60973, + "real_time": 1.1486110065110792e+04, + "cpu_time": 1.4469693963557680e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2822696153353588e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 60973, + "real_time": 1.1484426729864042e+04, + "cpu_time": 1.4467793809719048e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2826041400771198e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 60973, + "real_time": 1.1489720040017941e+04, + "cpu_time": 1.4491088625293195e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2815525451183285e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 60973, + "real_time": 1.1548099060239902e+04, + "cpu_time": 1.4452689486658005e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2700186293219601e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 60973, + "real_time": 1.1563139110753991e+04, + "cpu_time": 1.4449313147458734e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2670660405373825e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1514299001197334e+04, + "cpu_time": 1.4466115806537333e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2767021940780304e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1489720040017941e+04, + "cpu_time": 1.4467793809719048e+06, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2815525451183285e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8140838800137111e+01, + "cpu_time": 1.6597108155732410e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5312771849913478e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.3124759741058464e-03, + "cpu_time": 1.1473092278323992e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3079764251034167e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 36722, + "real_time": 1.9045029328467823e+04, + "cpu_time": 2.6227002680954197e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7528862831222488e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 36722, + "real_time": 1.9055588965742605e+04, + "cpu_time": 2.6210654847502890e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7513607736950272e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 36722, + "real_time": 1.9000430041936383e+04, + "cpu_time": 2.6227831526332912e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7593480718216862e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 36722, + "real_time": 1.9059863760143588e+04, + "cpu_time": 2.6227436187571473e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7507436915490852e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 36722, + "real_time": 1.9062880425902626e+04, + "cpu_time": 2.6230475547628067e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7503083914202072e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9044758504438607e+04, + "cpu_time": 2.6224680157997911e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7529294423216507e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9055588965742609e+04, + "cpu_time": 2.6227436187571478e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7513607736950272e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5685392798621557e+01, + "cpu_time": 7.9563938095902802e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7184944169768497e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3486856655407456e-03, + "cpu_time": 3.0339335929570018e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3507409088701166e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 20751, + "real_time": 3.3799569562912737e+04, + "cpu_time": 4.9675487230013097e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1023353656863464e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 20751, + "real_time": 3.3681486386198390e+04, + "cpu_time": 4.9704752362295892e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1132117744948261e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 20751, + "real_time": 3.3711679051612082e+04, + "cpu_time": 4.9679395265288418e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1104235371802330e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 20751, + "real_time": 3.3821257336995928e+04, + "cpu_time": 4.9720685589610180e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1003460029648224e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 20751, + "real_time": 3.3841316225724120e+04, + "cpu_time": 4.9689859451592593e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.0985083233935680e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3771061712688650e+04, + "cpu_time": 4.9694035979760047e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1049650007439590e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3799569562912737e+04, + "cpu_time": 4.9689859451592593e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.1023353656863464e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0388270293782014e+01, + "cpu_time": 1.8714249810563063e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4757953325454630e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0842776840307438e-03, + "cpu_time": 3.7658945266963657e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0856258704989726e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 11000, + "real_time": 6.3538785545454739e+04, + "cpu_time": 9.8952918783636708e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.3005855903552444e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 11000, + "real_time": 6.3433371363636506e+04, + "cpu_time": 9.8968007948181536e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.3060705349837402e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 11000, + "real_time": 6.3345416363636366e+04, + "cpu_time": 9.6205524379999842e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.3106610081481392e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 11000, + "real_time": 6.3293728272727749e+04, + "cpu_time": 9.6081091512727328e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.3133646211573051e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 11000, + "real_time": 6.3322588999999738e+04, + "cpu_time": 9.5921063742726725e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.3118544789759125e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3386778109091021e+04, + "cpu_time": 9.7225721273454428e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.3085072467240688e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3345416363636366e+04, + "cpu_time": 9.6205524379999824e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.3106610081481392e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.9707697597188115e+01, + "cpu_time": 1.5868116113100166e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1992052788796097e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5730046639314500e-03, + "cpu_time": 1.6320903465935659e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5714655858854844e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5699, + "real_time": 1.2286242410949325e+05, + "cpu_time": 1.9558165483418122e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4138216223555023e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5699, + "real_time": 1.2301683154939457e+05, + "cpu_time": 1.9562463860326473e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4095366846738155e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5699, + "real_time": 1.2319794472714519e+05, + "cpu_time": 1.9567679844183128e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4045243281366486e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5699, + "real_time": 1.2317852324969326e+05, + "cpu_time": 1.7564757709247265e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4050611172678146e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5699, + "real_time": 1.2294628794525375e+05, + "cpu_time": 1.7900312100017559e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4114929942965538e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2304040231619604e+05, + "cpu_time": 1.8830675799438510e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4088873493460670e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2301683154939460e+05, + "cpu_time": 1.9558165483418122e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4095366846738155e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4576208478366695e+02, + "cpu_time": 1.0094621330934910e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0383808048228435e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1846684669404729e-03, + "cpu_time": 5.3607323701233861e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1846624399593415e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2904, + "real_time": 2.4122196039944890e+05, + "cpu_time": 3.8492433592975341e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4775473949838455e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2904, + "real_time": 2.4219956473829222e+05, + "cpu_time": 3.8487878238980830e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4635107660347275e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2904, + "real_time": 2.4237050826446319e+05, + "cpu_time": 3.8493387943870619e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4610679575118729e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2904, + "real_time": 2.4179182885674963e+05, + "cpu_time": 3.8481144676652722e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4693513174797394e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2904, + "real_time": 2.4128205681818243e+05, + "cpu_time": 3.8484196019627966e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4766812379758591e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4177318381542727e+05, + "cpu_time": 3.8487808094421491e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4696317347972099e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4179182885674963e+05, + "cpu_time": 3.8487878238980830e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4693513174797394e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2057607706478632e+02, + "cpu_time": 5.2436413571268877e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.4705479229458392e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1531588774634357e-03, + "cpu_time": 1.3624162083386901e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1531241624357781e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1454, + "real_time": 6.0477286726272409e+05, + "cpu_time": 7.6611305074278250e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.7741350361724609e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1454, + "real_time": 8.0643370976616198e+05, + "cpu_time": 7.6834109156121135e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.0804209691165829e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1454, + "real_time": 5.2867635557083902e+05, + "cpu_time": 7.6538583692572221e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.1734379310163738e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1454, + "real_time": 4.8063188789546129e+05, + "cpu_time": 7.6311425679504484e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4906581153951836e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1454, + "real_time": 4.7745704951856908e+05, + "cpu_time": 7.6356301193260133e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.5138691567999367e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.7959437400275120e+05, + "cpu_time": 7.6530344959147245e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.0065042417001076e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2867635557083902e+05, + "cpu_time": 7.6538583692572206e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.1734379310163738e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3684689627717447e+05, + "cpu_time": 2.1042337546699081e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9818823595786095e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3610804799931495e-01, + "cpu_time": 2.7495417089693918e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9896470713761560e-01, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 596, + "real_time": 1.1415826275167784e+06, + "cpu_time": 1.5206439353523484e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.9392906997007389e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 596, + "real_time": 1.1990426946308720e+06, + "cpu_time": 1.5211229795973063e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.7984351308132366e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 596, + "real_time": 1.8717759546979845e+06, + "cpu_time": 1.3688409027013558e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.7926521556055614e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 596, + "real_time": 1.8951973976510044e+06, + "cpu_time": 1.5287307919463071e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.7704979988675014e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 596, + "real_time": 1.8754257315436241e+06, + "cpu_time": 1.5285901606879199e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.7891634648940239e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5966048812080531e+06, + "cpu_time": 1.4935857540570477e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.2180078899762131e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8717759546979847e+06, + "cpu_time": 1.5211229795973063e+08, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.7926521556055614e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8978154801551846e+05, + "cpu_time": 6.9843045604118761e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9628953073066216e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_coarse/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4413150216639365e-01, + "cpu_time": 4.6761992349219406e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6884013056286155e-01, + "dst_numa": NaN, + "src_gpu": NaN + } \ No newline at end of file diff --git a/results/implicit_managed_HostWrGPU_fine.json b/results/implicit_managed_HostWrGPU_fine.json new file mode 100644 index 0000000..432a493 --- /dev/null +++ b/results/implicit_managed_HostWrGPU_fine.json @@ -0,0 +1,3194 @@ +{ + "context": { + "date": "2023-10-09T20:12:11-04:00", + "host_name": "frontier03063", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.91,7.45,10.79], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 233544, + "real_time": 3.0155929632101765e+03, + "cpu_time": 2.6972809620028769e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.3582734971100683e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 233544, + "real_time": 2.9664183365874792e+03, + "cpu_time": 2.7088305599801318e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.3807897387500556e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 233544, + "real_time": 2.9808448386597784e+03, + "cpu_time": 2.6969252331894636e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.3741070809447460e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 233544, + "real_time": 2.9866053548794248e+03, + "cpu_time": 2.7062360487959447e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.3714567253782225e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 233544, + "real_time": 2.9829072037815968e+03, + "cpu_time": 2.6893949123505631e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.3731570311028361e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9864737394236918e+03, + "cpu_time": 2.6997335432637960e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.3715568146571858e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9829072037815968e+03, + "cpu_time": 2.6972809620028769e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.3731570311028361e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7990069481296459e+01, + "cpu_time": 7.8394410007078682e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2275145249680355e+06, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.0238498814886793e-03, + "cpu_time": 2.9037832345597007e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9986684015160274e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 227769, + "real_time": 3.0671435401656199e+03, + "cpu_time": 2.8031843178834685e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.6708890186331635e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 227769, + "real_time": 3.0813743573531651e+03, + "cpu_time": 2.8090530895775981e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.6585539600052857e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 227769, + "real_time": 3.0396734279027723e+03, + "cpu_time": 2.8139615667628165e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.6950263553976865e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 227769, + "real_time": 3.0625020086139061e+03, + "cpu_time": 2.8236379295250919e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.6749370210887513e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 227769, + "real_time": 3.0577794871118745e+03, + "cpu_time": 2.8097895610465016e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.6790682698108768e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0616945642294677e+03, + "cpu_time": 2.8119252929590951e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.6756949249871526e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0625020086139061e+03, + "cpu_time": 2.8097895610465016e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 2.6749370210887513e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5148397906455733e+01, + "cpu_time": 7.5919005308506871e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3252994866610339e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.9477168896722096e-03, + "cpu_time": 2.6998941080904190e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9531038620458474e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 199950, + "real_time": 3.5001496674166970e+03, + "cpu_time": 3.0591875745936448e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6809426901142464e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 199950, + "real_time": 3.4977831007751383e+03, + "cpu_time": 3.0573928405101289e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6841097712345762e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 199950, + "real_time": 3.5062230357587937e+03, + "cpu_time": 3.0529968596649187e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6728345096433039e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 199950, + "real_time": 3.4981670017503625e+03, + "cpu_time": 3.0512837552888208e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6835957207880611e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 199950, + "real_time": 3.4824986896722912e+03, + "cpu_time": 3.0500063197799452e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.7046679582646914e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4969642990746565e+03, + "cpu_time": 3.0541734699674917e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6852301300089760e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4981670017503625e+03, + "cpu_time": 3.0529968596649187e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 4.6835957207880611e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7641537210511569e+00, + "cpu_time": 3.9561848826779419e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1766363177532809e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5062176709582848e-03, + "cpu_time": 1.2953373217271943e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5113735827337613e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 161987, + "real_time": 4.3060596158948683e+03, + "cpu_time": 3.5207858705945557e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.6097413698231583e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 161987, + "real_time": 4.3477965762683352e+03, + "cpu_time": 3.5224695971281629e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.5366911549768057e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 161987, + "real_time": 4.3337889151598747e+03, + "cpu_time": 3.5192590670856240e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.5610512282625055e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 161987, + "real_time": 4.3352964929284335e+03, + "cpu_time": 3.5243725380431797e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.5584219103468208e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 161987, + "real_time": 4.3113688814534125e+03, + "cpu_time": 3.5236120694870577e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.6003703002452269e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3268620963409849e+03, + "cpu_time": 3.5220998284677160e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.5732551927309036e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3337889151598747e+03, + "cpu_time": 3.5224695971281629e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 7.5610512282625055e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7536784303540877e+01, + "cpu_time": 2.0847794308193292e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0709778761248019e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.0530028258517594e-03, + "cpu_time": 5.9191378221846406e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0550302320096678e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 130995, + "real_time": 5.3446849498071715e+03, + "cpu_time": 4.4082182826062012e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2261901424585268e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 130995, + "real_time": 5.3483000877894810e+03, + "cpu_time": 4.4037548947669729e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2253613096546877e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 130995, + "real_time": 5.3524073285238683e+03, + "cpu_time": 4.4054883907782723e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2244210124059088e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 130995, + "real_time": 5.3427888392685318e+03, + "cpu_time": 4.4108085467384203e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2266253069618298e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 130995, + "real_time": 5.3464292911941393e+03, + "cpu_time": 4.4104057435779908e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2257900821383980e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3469220993166382e+03, + "cpu_time": 4.4077351716935710e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2256775707238701e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3464292911941393e+03, + "cpu_time": 4.4082182826062012e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 1.2257900821383980e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6850416588673220e+00, + "cpu_time": 3.0700819779664403e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4447505144438688e+06, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.8918932993960904e-04, + "cpu_time": 6.9652142390098080e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8898629755103562e-04, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 94802, + "real_time": 7.3286408092655684e+03, + "cpu_time": 6.2206303115968022e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.7884898907077866e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 94802, + "real_time": 7.2996512731797684e+03, + "cpu_time": 6.2101174215733993e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.7955926262064339e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 94802, + "real_time": 7.3895636062528902e+03, + "cpu_time": 6.2005402227801015e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.7737447971770576e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 94802, + "real_time": 7.3284124174593690e+03, + "cpu_time": 6.2095052914495463e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.7885456294426231e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 94802, + "real_time": 7.3279987236554716e+03, + "cpu_time": 6.2068493947385217e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.7886465997447189e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.3348533659626146e+03, + "cpu_time": 6.2095285284276749e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.7870039086557243e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.3284124174593690e+03, + "cpu_time": 6.2095052914495463e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.7885456294426231e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3013124391397554e+01, + "cpu_time": 7.2726253920306044e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.0133261824422449e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.5008567648529892e-03, + "cpu_time": 1.1712041194006911e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4842242054581056e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 58081, + "real_time": 1.2068334446721001e+04, + "cpu_time": 9.7590792515624722e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.1721638653395557e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 58081, + "real_time": 1.2051047760885724e+04, + "cpu_time": 9.7610227699247876e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.1752797366785397e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 58081, + "real_time": 1.2053804411081072e+04, + "cpu_time": 9.7689307697870443e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.1747822601054550e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 58081, + "real_time": 1.2111342952084315e+04, + "cpu_time": 9.7622293820699118e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.1644503094092144e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 58081, + "real_time": 1.2095061500318488e+04, + "cpu_time": 9.7584571326251607e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.1673639277741348e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2075918214218120e+04, + "cpu_time": 9.7619438611938769e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.1708080198613800e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2068334446721001e+04, + "cpu_time": 9.7610227699247876e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.1721638653395557e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6393421165064346e+01, + "cpu_time": 4.1865883303551635e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7410812656733066e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1856243721482707e-03, + "cpu_time": 4.2886830634192448e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1840168371849182e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 34655, + "real_time": 2.0139374577983006e+04, + "cpu_time": 1.6928733145577945e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6032983197659378e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 34655, + "real_time": 2.0148679007358183e+04, + "cpu_time": 1.6895092370220756e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6020961463951706e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 34655, + "real_time": 2.0186394661665247e+04, + "cpu_time": 1.6911347768287421e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.5972344680035583e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 34655, + "real_time": 2.0168800779108376e+04, + "cpu_time": 1.8541654776511262e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.5995001177416443e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 34655, + "real_time": 2.0073006030875807e+04, + "cpu_time": 1.6918473509450241e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6119057563852322e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0143251011398126e+04, + "cpu_time": 1.7239060314009530e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6028069616583088e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0148679007358187e+04, + "cpu_time": 1.6918473509450243e+06, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6020961463951706e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3281093889144401e+01, + "cpu_time": 7.2827547045097075e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6018447590971313e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1486647743531396e-03, + "cpu_time": 4.2245659402857891e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1522321254005198e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 19553, + "real_time": 3.5784991305681935e+04, + "cpu_time": 3.1320028194139125e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.9302116941789146e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 19553, + "real_time": 3.5765395796041747e+04, + "cpu_time": 3.1700023139671576e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.9318171284324181e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 19553, + "real_time": 3.5684173579502276e+04, + "cpu_time": 3.1451249068685090e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.9384903580962395e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 19553, + "real_time": 3.5684550452616015e+04, + "cpu_time": 3.1912201641691751e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.9384593239933319e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 19553, + "real_time": 3.5686891116452483e+04, + "cpu_time": 3.1382352643072740e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.9382665936865044e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5721200450058896e+04, + "cpu_time": 3.1553170937452060e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.9354490196774818e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5686891116452476e+04, + "cpu_time": 3.1451249068685086e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.9382665936865044e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9784143298018499e+01, + "cpu_time": 2.4717964705327344e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0887225125336155e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3936861771378800e-03, + "cpu_time": 7.8337498168807949e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3928780520885502e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10473, + "real_time": 6.7095012603838288e+04, + "cpu_time": 6.3003736228396716e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1256451390546852e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10473, + "real_time": 6.7100081638499090e+04, + "cpu_time": 6.3291086062255101e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1254090141028175e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10473, + "real_time": 6.7299413921512591e+04, + "cpu_time": 6.0527408213501293e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1161519511087971e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10473, + "real_time": 6.7278811897259773e+04, + "cpu_time": 6.0274090688436842e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1171061748274063e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10473, + "real_time": 6.7238112097775287e+04, + "cpu_time": 6.1543367693115389e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1189929856305237e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7202286431777029e+04, + "cpu_time": 6.1727937777141081e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1206610529448463e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7238112097775287e+04, + "cpu_time": 6.1543367693115380e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.1189929856305237e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8141311887165585e+01, + "cpu_time": 1.3838445501483406e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5589432946015023e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4603865001944166e-03, + "cpu_time": 2.2418447788495575e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4608902464109022e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5414, + "real_time": 1.2919385592907260e+05, + "cpu_time": 1.2399680655153351e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.2465197124410255e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5414, + "real_time": 1.2919776172885112e+05, + "cpu_time": 1.2335255346693756e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.2464215663446518e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5414, + "real_time": 1.2916756206132255e+05, + "cpu_time": 1.2279013108976755e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.2471805870337215e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5414, + "real_time": 1.2923549261174719e+05, + "cpu_time": 1.2278606989471745e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.2454737589778397e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5414, + "real_time": 1.2912523328407796e+05, + "cpu_time": 1.2277392034355346e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.2482450511995991e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2918398112301431e+05, + "cpu_time": 1.2313989626930190e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.2467681351993679e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2919385592907260e+05, + "cpu_time": 1.2279013108976755e+07, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.2465197124410255e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0809755325431794e+01, + "cpu_time": 5.3874518046808233e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0257474122578131e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1590414671127886e-04, + "cpu_time": 4.3750660573066320e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1592875423942985e-04, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2749, + "real_time": 2.5397415896689752e+05, + "cpu_time": 2.4478735709712744e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3029376036218529e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2749, + "real_time": 2.5364987740996684e+05, + "cpu_time": 2.4125952863950539e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3071602815883644e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2749, + "real_time": 2.5357186395052742e+05, + "cpu_time": 2.4432672073845122e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3081777565182236e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2749, + "real_time": 2.5350915605674818e+05, + "cpu_time": 2.4321586415423706e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3089960656577652e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2749, + "real_time": 2.5352534521644190e+05, + "cpu_time": 2.4168985393233813e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3087847658144016e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5364608032011637e+05, + "cpu_time": 2.4305586491233185e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3072112946401215e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5357186395052742e+05, + "cpu_time": 2.4321586415423706e+07, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3081777565182236e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9136568723157765e+02, + "cpu_time": 1.5597715759635513e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4931057408807993e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5445946962816402e-04, + "cpu_time": 6.4173377446627225e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5383926782098443e-04, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1264, + "real_time": 5.1033588212025363e+05, + "cpu_time": 4.7252602200158454e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.2874850834115326e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1264, + "real_time": 5.2193674604430463e+05, + "cpu_time": 4.7662219351265542e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.2144155641756374e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1264, + "real_time": 5.0765259810126584e+05, + "cpu_time": 4.8209250250000089e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3048616441146046e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1264, + "real_time": 5.1406913765822758e+05, + "cpu_time": 4.8777460011866808e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.2636108202150276e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1264, + "real_time": 7.1706114952531713e+05, + "cpu_time": 4.7884163906645238e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.3397190059880173e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.5421110268987378e+05, + "cpu_time": 4.7957139143987231e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.0820184235809647e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1406913765822758e+05, + "cpu_time": 4.7884163906645238e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.2636108202150276e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1194679857644951e+04, + "cpu_time": 5.7557834190465196e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1634987569095230e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6454863393214225e-01, + "cpu_time": 1.2001932395853033e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3509000222237472e-01, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 593, + "real_time": 1.1514905564924108e+06, + "cpu_time": 9.4004740807756484e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.9139997554310074e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 593, + "real_time": 1.1578490876897140e+06, + "cpu_time": 9.4034433478921562e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.8979970150472736e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 593, + "real_time": 1.2039654991568308e+06, + "cpu_time": 9.4025399983135730e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.7869928186064354e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 593, + "real_time": 1.9019429021922413e+06, + "cpu_time": 9.4730252529510960e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.7642186819238407e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 593, + "real_time": 1.1352586543001689e+06, + "cpu_time": 9.3864301204047531e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.9556640570764606e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3101013399662734e+06, + "cpu_time": 9.4131825600674465e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.6637744656170036e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1578490876897140e+06, + "cpu_time": 9.4025399983135730e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.8979970150472736e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3183178915333422e+05, + "cpu_time": 3.4155756249001145e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0671880822003489e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5328711530199394e-01, + "cpu_time": 3.6285024784175036e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9022586737750144e-01, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 445, + "real_time": 1.5504821573033708e+06, + "cpu_time": 1.8873190292359531e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.3282577412381882e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 445, + "real_time": 1.7198073842696643e+06, + "cpu_time": 1.8565847643820214e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.9021151213685791e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 445, + "real_time": 1.5641868988764065e+06, + "cpu_time": 1.8845961093033561e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.2903353843588600e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 445, + "real_time": 1.5633559573033708e+06, + "cpu_time": 1.8861822754606897e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.2926157466886765e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 445, + "real_time": 1.5737859797752826e+06, + "cpu_time": 1.8741907801123601e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.2641671016526871e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5943236755056190e+06, + "cpu_time": 1.8777745916988760e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.2154982190613983e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5641868988764063e+06, + "cpu_time": 1.8845961093033561e+08, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.2903353843588600e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0635086369353899e+04, + "cpu_time": 1.2943830806310396e+06, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7666355755424602e+09, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.4304106784936820e-02, + "cpu_time": 6.8931760305691132e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1908108691736334e-02, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 210, + "real_time": 3.3365120190476188e+06, + "cpu_time": 3.8571235515238148e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 4.0226957743227730e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 210, + "real_time": 3.3691591999999988e+06, + "cpu_time": 3.7573143566666353e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.9837158184748299e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 210, + "real_time": 3.3978972190476190e+06, + "cpu_time": 4.0290710471428519e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.9500231863287277e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 210, + "real_time": 3.4610021571428576e+06, + "cpu_time": 3.9121018689523959e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.8780018591724899e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 210, + "real_time": 3.4615118333333326e+06, + "cpu_time": 3.7853431794285691e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.8774308586070129e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4052164857142856e+06, + "cpu_time": 3.8681908007428539e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.9423734993811668e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3978972190476186e+06, + "cpu_time": 3.8571235515238142e+08, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.9500231863287277e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.5576997395130515e+04, + "cpu_time": 1.0850165183848687e+07, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4382892844505262e+08, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6321134831893826e-02, + "cpu_time": 2.8049715597702685e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6330997774465415e-02, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 69, + "real_time": 9.3272018260869551e+06, + "cpu_time": 8.4955534697100902e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8779848555353584e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 69, + "real_time": 9.2841959420289826e+06, + "cpu_time": 7.1320964498551369e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8913161427885132e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 69, + "real_time": 9.2858304347826056e+06, + "cpu_time": 7.4167006176811922e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8908072130469013e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 69, + "real_time": 9.3081413188405782e+06, + "cpu_time": 7.1584344262319589e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8838781750837910e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 69, + "real_time": 9.2382806086956505e+06, + "cpu_time": 7.4915016449275470e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.9056863216227882e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.2887300260869544e+06, + "cpu_time": 7.5388573216811860e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8899345416154709e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.2858304347826056e+06, + "cpu_time": 7.4167006176811934e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.8908072130469013e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3282829449290366e+04, + "cpu_time": 5.5736171521214902e+07, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0371397249086316e+08, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.5831410059090018e-03, + "cpu_time": 7.3931856172581856e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5888000574880543e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 30, + "real_time": 2.2830802933333337e+07, + "cpu_time": 1.6559376866333272e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.3515200650966148e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 30, + "real_time": 2.2982634666666668e+07, + "cpu_time": 1.3811085939666555e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.3359850590962124e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 30, + "real_time": 2.2703545000000000e+07, + "cpu_time": 1.5044411888666825e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.3647008077372940e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 30, + "real_time": 2.2831941333333332e+07, + "cpu_time": 1.5047648200999901e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.3514028183674377e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 30, + "real_time": 2.2634808399999999e+07, + "cpu_time": 1.5055254827666736e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.3718818490197605e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2796746466666669e+07, + "cpu_time": 1.5103555544666657e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.3550981198634640e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2830802933333337e+07, + "cpu_time": 1.5047648200999901e+09, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.3515200650966148e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3403674429560150e+05, + "cpu_time": 9.7453193416788816e+07, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3835389403147164e+08, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.8796435926323784e-03, + "cpu_time": 6.4523345598051132e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8746551943870884e-03, + "dst_numa": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 13, + "real_time": 5.0348478692307696e+07, + "cpu_time": 2.6770318322307782e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.1326201940716190e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 13, + "real_time": 5.0364640923076913e+07, + "cpu_time": 3.2608420600769162e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.1319358270417351e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 13, + "real_time": 5.0375775384615406e+07, + "cpu_time": 3.2624082286923184e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.1314646093326778e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 13, + "real_time": 5.0059383153846152e+07, + "cpu_time": 3.2635410371538873e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.1449361864889507e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 13, + "real_time": 5.0126751153846145e+07, + "cpu_time": 3.2641150416923037e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.1420534929633347e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0255005861538470e+07, + "cpu_time": 3.1455876399692411e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.1366020619796638e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0348478692307696e+07, + "cpu_time": 3.2624082286923184e+09, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.1326201940716190e+10, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5004983199249278e+05, + "cpu_time": 2.6193363908440846e+08, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.3874519827809900e+07, + "dst_numa": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_managed_HostWrGPU_fine/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9857688685960343e-03, + "cpu_time": 8.3270176852223945e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9895374980883015e-03, + "dst_numa": NaN, + "src_gpu": NaN + } + ] +} diff --git a/results/implicit_mapped_GPURdHost.json b/results/implicit_mapped_GPURdHost.json new file mode 100644 index 0000000..ddd793f --- /dev/null +++ b/results/implicit_mapped_GPURdHost.json @@ -0,0 +1,3194 @@ +{ + "context": { + "date": "2023-10-09T17:38:31-04:00", + "host_name": "frontier08320", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1790, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.63,11.72,27.46], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52486, + "real_time": 1.3343102133880950e+04, + "cpu_time": 2.0306559463475980e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0697509161676824e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52486, + "real_time": 1.3339251929512149e+04, + "cpu_time": 2.0319817951453722e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0706369604864353e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52486, + "real_time": 1.3346214239978441e+04, + "cpu_time": 2.0285791163357859e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0690351034006906e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52486, + "real_time": 1.3339745220734190e+04, + "cpu_time": 2.0321902412071791e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0705234112219161e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52486, + "real_time": 1.3337333569954037e+04, + "cpu_time": 2.0307071180886320e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0710786219123679e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3341129418811954e+04, + "cpu_time": 2.0308228434249137e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0702050026378185e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3339745220734190e+04, + "cpu_time": 2.0307071180886320e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0705234112219161e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5212373440512912e+00, + "cpu_time": 1.4394881052279132e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1025319993197190e+04, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6393847428584965e-04, + "cpu_time": 7.0882012672275510e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6390850097496072e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52394, + "real_time": 1.3363867479897204e+04, + "cpu_time": 2.0549271653242744e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1299620131095564e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52394, + "real_time": 1.3363975336102118e+04, + "cpu_time": 2.0583252376226294e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1299125402227569e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52394, + "real_time": 1.3359705347024312e+04, + "cpu_time": 2.0437514925373140e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1318717645405650e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52394, + "real_time": 1.3374915532104196e+04, + "cpu_time": 2.0561064492117413e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1248984940028262e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52394, + "real_time": 1.3375717797981242e+04, + "cpu_time": 2.0551304786807646e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1245311270221281e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3367636298621816e+04, + "cpu_time": 2.0536481646753451e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1282351877795672e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3363975336102119e+04, + "cpu_time": 2.0551304786807646e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1299125402227569e+08, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.2250347130620920e+00, + "cpu_time": 5.6941942162246541e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3118619717615045e+05, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.4048708026317141e-04, + "cpu_time": 2.7727214009537178e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4042670855155054e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52192, + "real_time": 1.3413764894198763e+04, + "cpu_time": 2.0581944991569599e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2214318745877092e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52192, + "real_time": 1.3413961494987139e+04, + "cpu_time": 2.0607792822654785e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2214139727568755e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52192, + "real_time": 1.3415167311312191e+04, + "cpu_time": 2.0660646440067452e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2213041865071912e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52192, + "real_time": 1.3418416827962776e+04, + "cpu_time": 2.0646030196198637e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2210084252157986e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52192, + "real_time": 1.3416950300447232e+04, + "cpu_time": 2.0610562059319480e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2211418864280856e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3415652165781619e+04, + "cpu_time": 2.0621395301961995e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2212600690991321e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3415167311312191e+04, + "cpu_time": 2.0610562059319484e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2213041865071912e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9989845020697405e+00, + "cpu_time": 3.1647446326224479e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8196263000957092e+05, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4900390062053134e-04, + "cpu_time": 1.5346898627763286e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4899580737442474e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51044, + "real_time": 1.3711991713686977e+04, + "cpu_time": 2.0764886470496058e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3897330660790710e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51044, + "real_time": 1.3712509132962894e+04, + "cpu_time": 2.0746839099600333e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3896428933805013e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51044, + "real_time": 1.3711090591453749e+04, + "cpu_time": 2.0773696360003152e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3898901244533091e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51044, + "real_time": 1.3711786539226337e+04, + "cpu_time": 2.0766161292218458e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3897688245261197e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51044, + "real_time": 1.3719052868123958e+04, + "cpu_time": 2.0775694283363395e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3885030777989073e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3713286169090783e+04, + "cpu_time": 2.0765455501136279e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3895075972475815e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3711991713686977e+04, + "cpu_time": 2.0766161292218458e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.3897330660790710e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2635742906672682e+00, + "cpu_time": 1.1403253473016811e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6849963039565820e+05, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3798630397017735e-04, + "cpu_time": 5.4914535693150715e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3791497087119530e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49230, + "real_time": 1.4220863946114063e+04, + "cpu_time": 2.1574719337802155e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6084401234925051e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49230, + "real_time": 1.4220461022558291e+04, + "cpu_time": 2.1584968697948421e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6085706993632994e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49230, + "real_time": 1.4220617288231986e+04, + "cpu_time": 2.1591793398334350e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6085200572997017e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49230, + "real_time": 1.4217162202084741e+04, + "cpu_time": 2.1574561425959779e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6096400300187969e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49230, + "real_time": 1.4218575769694855e+04, + "cpu_time": 2.1582937659963409e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6091817536100855e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4219536045736784e+04, + "cpu_time": 2.1581796104001620e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6088705327568769e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4220461022558293e+04, + "cpu_time": 2.1582937659963409e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.6085706993632994e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6082016901689915e+00, + "cpu_time": 7.3098492330453499e+00, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2128920311090274e+05, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1309804236905132e-04, + "cpu_time": 3.3870439688242553e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1310562954761163e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 45881, + "real_time": 1.5265313958077182e+04, + "cpu_time": 2.3268341797258166e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.5862629723804131e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 45881, + "real_time": 1.5262242516505810e+04, + "cpu_time": 2.3286020814716197e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.5879909101331768e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 45881, + "real_time": 1.5260014907651184e+04, + "cpu_time": 2.3282706893921168e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.5892445579644947e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 45881, + "real_time": 1.5259344871797617e+04, + "cpu_time": 2.3262757786447535e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.5896217105786638e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 45881, + "real_time": 1.5262358140061169e+04, + "cpu_time": 2.3259279440291117e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.5879258498041439e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5261854878818593e+04, + "cpu_time": 2.3271821346526838e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.5882092001721792e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5262242516505808e+04, + "cpu_time": 2.3268341797258163e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.5879909101331768e+09, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3481052539824732e+00, + "cpu_time": 1.1954861306656847e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3212581058975570e+06, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5385451327029246e-04, + "cpu_time": 5.1370544353379665e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5384558935418899e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 37054, + "real_time": 1.8885115725230247e+04, + "cpu_time": 2.9050974469692912e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3880984570816231e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 37054, + "real_time": 1.8891096366617265e+04, + "cpu_time": 2.9036633345927610e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3876590056638456e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 37054, + "real_time": 1.8892084694927587e+04, + "cpu_time": 2.9066685944837238e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3875864110983162e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 37054, + "real_time": 1.8897374197020203e+04, + "cpu_time": 2.9069922950288692e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3871980163325321e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 37054, + "real_time": 1.8891972905083647e+04, + "cpu_time": 2.9059558347277030e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3875946218907587e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8891528777775791e+04, + "cpu_time": 2.9056755011604699e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3876273024134153e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8891972905083647e+04, + "cpu_time": 2.9059558347277034e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3875946218907587e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3585961411553553e+00, + "cpu_time": 1.3399373374880156e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2016342386974813e+06, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3071696274167379e-04, + "cpu_time": 4.6114486526553666e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3072724449346555e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 23894, + "real_time": 2.9303606337323636e+04, + "cpu_time": 4.4703669749727938e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7891586242483093e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 23894, + "real_time": 2.9307499300168220e+04, + "cpu_time": 4.4655646354733544e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7889209673955044e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 23894, + "real_time": 2.9308422158664038e+04, + "cpu_time": 4.4665849376412443e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7888646381634438e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 23894, + "real_time": 2.9334530341879366e+04, + "cpu_time": 4.4699899472671073e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7872725211199360e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 23894, + "real_time": 2.9326694006645113e+04, + "cpu_time": 4.4658923662844332e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7877500951222187e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9316150428936078e+04, + "cpu_time": 4.4676797723277865e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7883933692098827e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9308422158664041e+04, + "cpu_time": 4.4665849376412443e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.7888646381634438e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3609837028950931e+01, + "cpu_time": 2.3143710582058052e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.3008607035559872e+06, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.6424366193446539e-04, + "cpu_time": 5.1802527847692032e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6415183854228508e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 14029, + "real_time": 4.9889685089186736e+04, + "cpu_time": 7.6843566825860747e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.1017891737049110e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 14029, + "real_time": 4.9892494107754537e+04, + "cpu_time": 7.6831654786513682e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.1016708399771603e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 14029, + "real_time": 4.9904649665934310e+04, + "cpu_time": 7.6819295815810139e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.1011589241067738e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 14029, + "real_time": 4.9900579813739889e+04, + "cpu_time": 7.6789093734407128e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.1013302929824467e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 14029, + "real_time": 4.9887200143243979e+04, + "cpu_time": 7.6854342647373516e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.1018938665412441e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9894921763971892e+04, + "cpu_time": 7.6827590761993051e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.1015686194625072e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9892494107754537e+04, + "cpu_time": 7.6831654786513682e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.1016708399771603e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4092055439071585e+00, + "cpu_time": 2.5190482096878746e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1206115831355881e+06, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4849618522214410e-04, + "cpu_time": 3.2788327535764132e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4848963551490930e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 7690, + "real_time": 9.1041077037336756e+04, + "cpu_time": 1.3212632392717735e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3035228363345695e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 7690, + "real_time": 9.1048697564055750e+04, + "cpu_time": 1.3205199609882943e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3033300377796009e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 7690, + "real_time": 9.1058370041391798e+04, + "cpu_time": 1.3210769479843948e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3030853715553127e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 7690, + "real_time": 9.1019157768437290e+04, + "cpu_time": 1.3211910754226355e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3040775715980415e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 7690, + "real_time": 9.1012013161254290e+04, + "cpu_time": 1.3210559596879064e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3042584458430611e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1035863114495165e+04, + "cpu_time": 1.3210214366710008e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3036548526221176e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1041077037336741e+04, + "cpu_time": 1.3210769479843951e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3035228363345695e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9661883843241462e+01, + "cpu_time": 2.9282973257362620e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9755848099454604e+06, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1597954004690266e-04, + "cpu_time": 2.2166917541592863e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1598655737347280e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4041, + "real_time": 1.7318660028281016e+05, + "cpu_time": 2.4036481935164420e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.4218409467884857e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4041, + "real_time": 1.7320152540190850e+05, + "cpu_time": 2.4408785919326896e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.4216322519488522e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4041, + "real_time": 1.7319768411041104e+05, + "cpu_time": 2.4034033407572468e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.4216859604925156e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4041, + "real_time": 1.7318921038361007e+05, + "cpu_time": 2.4417392279138879e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.4218044476960857e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4041, + "real_time": 1.7319070819017000e+05, + "cpu_time": 2.4063223509032457e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.4217835031856873e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7319314567378192e+05, + "cpu_time": 2.4191983410047024e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.4217494220223255e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7319070819016997e+05, + "cpu_time": 2.4063223509032457e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 2.4217835031856873e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2267008687813981e+00, + "cpu_time": 2.0218834683096034e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7066527722196432e+05, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.5952351604662836e-05, + "cpu_time": 8.3576589568505871e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5951914318819144e-05, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2076, + "real_time": 3.3722065106753854e+05, + "cpu_time": 4.5794066184971726e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4875724465403305e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2076, + "real_time": 3.3731969929538463e+05, + "cpu_time": 4.6686526107899164e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4868420129398525e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2076, + "real_time": 3.3723389918443537e+05, + "cpu_time": 4.5813599421964929e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4874747231185726e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2076, + "real_time": 3.3719461536401010e+05, + "cpu_time": 4.6696037186898006e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4877645187021404e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2076, + "real_time": 3.3721462737372762e+05, + "cpu_time": 4.5788245520231297e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4876168822602966e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3723669845701929e+05, + "cpu_time": 4.6155694884393021e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4874541167122391e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3722065106753854e+05, + "cpu_time": 4.5813599421964929e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 2.4875724465403305e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8509311510500360e+01, + "cpu_time": 4.8902332475231415e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5774977588252937e+06, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4384351327257125e-04, + "cpu_time": 1.0595080974886836e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4382165824846675e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1048, + "real_time": 6.6790240000309562e+05, + "cpu_time": 9.2030118893129460e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.5119262934108696e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1048, + "real_time": 6.6809219269096164e+05, + "cpu_time": 9.2119584351145430e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.5112127014123947e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1048, + "real_time": 6.6819737410085730e+05, + "cpu_time": 8.9087389312977856e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.5108174096876438e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1048, + "real_time": 6.6832454395453795e+05, + "cpu_time": 8.9286794847328786e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.5103396473706722e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1048, + "real_time": 6.6820438273645635e+05, + "cpu_time": 8.9291111164122855e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.5107910743256878e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6814417869718187e+05, + "cpu_time": 9.0362999713740894e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.5110174252414539e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6819737410085730e+05, + "cpu_time": 8.9291111164122855e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 2.5108174096876438e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5822877788437012e+02, + "cpu_time": 1.5651852780481300e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9472639187310329e+06, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3681831396466156e-04, + "cpu_time": 1.7321085875927632e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3684678007199241e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 526, + "real_time": 1.3327013887996173e+06, + "cpu_time": 2.1760614144486729e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.5177757209529846e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 526, + "real_time": 1.3334095022577383e+06, + "cpu_time": 2.1785010304182530e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.5164386441813564e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 526, + "real_time": 1.3342755288948012e+06, + "cpu_time": 2.1845527642585495e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.5148053211913136e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 526, + "real_time": 1.3351263093298208e+06, + "cpu_time": 2.1861972167300521e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.5132028157577808e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 526, + "real_time": 1.3319494035856326e+06, + "cpu_time": 2.1842759315589219e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.5191971939527771e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3334924265735219e+06, + "cpu_time": 2.1819176714828899e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.5162839392072426e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3334095022577380e+06, + "cpu_time": 2.1842759315589219e+06, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 2.5164386441813564e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2546273599501972e+03, + "cpu_time": 4.3814607912450028e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3672786364174370e+07, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.4085825682117081e-04, + "cpu_time": 2.0080779621108455e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.4078358945582675e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 264, + "real_time": 2.6459336919343155e+06, + "cpu_time": 5.5874611628787816e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5363018054673893e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 264, + "real_time": 2.6506628112684032e+06, + "cpu_time": 5.5821029696969744e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5317767207020519e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 264, + "real_time": 2.6500707026571035e+06, + "cpu_time": 5.5802771174242049e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5323423987410236e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 264, + "real_time": 2.6508046465812987e+06, + "cpu_time": 5.5665409696969576e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5316412541584026e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 264, + "real_time": 2.6513289194554090e+06, + "cpu_time": 5.5961295871212175e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5311406482823097e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6497601543793064e+06, + "cpu_time": 5.5825023613636279e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5326405654702358e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6506628112684032e+06, + "cpu_time": 5.5821029696969744e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 2.5317767207020519e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1854651882269236e+03, + "cpu_time": 1.0841681632483782e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0909438152734760e+07, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.2477849348552010e-04, + "cpu_time": 1.9420827669539050e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2559832760368428e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 132, + "real_time": 5.2858977543303007e+06, + "cpu_time": 1.2190734575757572e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5391661783477833e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 132, + "real_time": 5.2881886702821106e+06, + "cpu_time": 1.2202865666666700e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5380661766903229e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 132, + "real_time": 5.2873208328629984e+06, + "cpu_time": 1.2205877606060619e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5384827636291416e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 132, + "real_time": 5.2884226474402985e+06, + "cpu_time": 1.2222263886363639e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5379538843205746e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 132, + "real_time": 5.2894445340798208e+06, + "cpu_time": 1.2195298522727255e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5374635679651608e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2878548877991065e+06, + "cpu_time": 1.2203408051515158e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5382265141905968e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2881886702821115e+06, + "cpu_time": 1.2202865666666698e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 2.5380661766903229e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3300679612408417e+03, + "cpu_time": 1.2127621714253697e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.3850501103248987e+06, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5153261378442216e-04, + "cpu_time": 9.9378973996923337e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5155556742582517e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 66, + "real_time": 1.0555863592096351e+07, + "cpu_time": 2.5612818257575732e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5429985302291107e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 66, + "real_time": 1.0564164230317781e+07, + "cpu_time": 2.5615035606060650e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5410004061620422e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 66, + "real_time": 1.0558644224974243e+07, + "cpu_time": 2.5611541333333321e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5423288282133102e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 66, + "real_time": 1.0555841861236276e+07, + "cpu_time": 2.5569548606060654e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5430037653913986e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 66, + "real_time": 1.0553475779791674e+07, + "cpu_time": 2.5571405878787860e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5435739049500046e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0557597937683266e+07, + "cpu_time": 2.5596069936363645e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5425810869891735e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0555863592096351e+07, + "cpu_time": 2.5611541333333321e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.5429985302291107e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1016722431615544e+03, + "cpu_time": 2.3405467906297825e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.8752085526858624e+06, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.8850430442339954e-04, + "cpu_time": 9.1441646957864835e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8839306259371667e-04, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1266342168956093e+07, + "cpu_time": 5.2715656666666783e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5245098933078701e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1188150244680319e+07, + "cpu_time": 5.2617559878787763e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5338262462755165e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1201551694310073e+07, + "cpu_time": 5.2618081515151709e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5322246208237755e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1234031935984436e+07, + "cpu_time": 5.2736683060606048e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5283512505704910e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1229285295262482e+07, + "cpu_time": 5.2784332454545319e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5289165628190407e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1223872267838679e+07, + "cpu_time": 5.2694462715151526e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5295657147593391e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1229285295262482e+07, + "cpu_time": 5.2715656666666776e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 2.5289165628190407e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0456867718878853e+04, + "cpu_time": 7.4257062225184316e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6290343861658849e+07, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4350287890222216e-03, + "cpu_time": 1.4092004813976928e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4346472064320924e-03, + "own_numa": NaN, + "rd_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 16, + "real_time": 4.2643473250791430e+07, + "cpu_time": 1.0680745068750052e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5179511473776875e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 16, + "real_time": 4.2610012926161289e+07, + "cpu_time": 1.0674474306250037e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5199284164984474e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 16, + "real_time": 4.2571332072839141e+07, + "cpu_time": 1.0588586868750039e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5222180554811817e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 16, + "real_time": 4.2586072348058224e+07, + "cpu_time": 1.0591637843750057e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5213450426333080e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 16, + "real_time": 4.2525032069534063e+07, + "cpu_time": 1.0585384862500025e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5249641722651493e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2587184533476830e+07, + "cpu_time": 1.0624165790000042e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5212813668511551e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2586072348058224e+07, + "cpu_time": 1.0591637843750057e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 2.5213450426333080e+10, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4173016220148464e+04, + "cpu_time": 4.8887751158700389e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6155942107193921e+07, + "own_numa": 0.0000000000000000e+00, + "rd_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPURdHost/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0372372981225148e-03, + "cpu_time": 4.6015613955041820e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0374067111700528e-03, + "own_numa": NaN, + "rd_gpu": NaN + } + ] +} diff --git a/results/implicit_mapped_GPUWrGPU.json b/results/implicit_mapped_GPUWrGPU.json new file mode 100644 index 0000000..41e096a --- /dev/null +++ b/results/implicit_mapped_GPUWrGPU.json @@ -0,0 +1,25272 @@ +{ + "context": { + "date": "2023-10-09T17:45:33-04:00", + "host_name": "frontier05673", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.29,3.2,19.1], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57046, + "real_time": 1.2267238004456958e+04, + "cpu_time": 2.8659642726922131e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3389749171833408e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57046, + "real_time": 1.2267363340961074e+04, + "cpu_time": 2.8685867457841036e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3389408026444769e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57046, + "real_time": 1.2266843972453978e+04, + "cpu_time": 2.8643220593906663e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3390821707668602e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57046, + "real_time": 1.2267951096180519e+04, + "cpu_time": 2.8658853626897599e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3387808346213901e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57046, + "real_time": 1.2270554779584207e+04, + "cpu_time": 2.8656289660975348e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3380723802439147e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2267990238727349e+04, + "cpu_time": 2.8660774813308562e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3387702210919970e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2267363340961074e+04, + "cpu_time": 2.8658853626897599e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3389408026444769e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4875471038461685e+00, + "cpu_time": 1.5515158688378810e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0478563462652673e+04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2125434361288528e-04, + "cpu_time": 5.4133772689126270e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2123794326107750e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57079, + "real_time": 1.2266389029302303e+04, + "cpu_time": 2.8631468263284209e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6784120252755022e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57079, + "real_time": 1.2262311124837492e+04, + "cpu_time": 2.8661623101315730e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6806329708981073e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57079, + "real_time": 1.2261666860868856e+04, + "cpu_time": 2.8634291911210770e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6809839909641123e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57079, + "real_time": 1.2265609271990548e+04, + "cpu_time": 2.8638970724784984e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6788365896401525e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57079, + "real_time": 1.2264726635322684e+04, + "cpu_time": 2.8624457453704552e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6793172351733136e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2264140584464374e+04, + "cpu_time": 2.8638162290860044e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6796365623902380e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2264726635322684e+04, + "cpu_time": 2.8634291911210770e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6793172351733136e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0629004981107664e+00, + "cpu_time": 1.4130424599038241e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1235892381115085e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6820587499820000e-04, + "cpu_time": 4.9341240738579116e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6821113358740043e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57092, + "real_time": 1.2259492711462217e+04, + "cpu_time": 2.8882859822742234e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3364337648882899e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57092, + "real_time": 1.2262444678003701e+04, + "cpu_time": 2.8930858999509579e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3361120421109438e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57092, + "real_time": 1.2255068018272501e+04, + "cpu_time": 2.8861759160653026e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3369162843952556e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57092, + "real_time": 1.2258835438477276e+04, + "cpu_time": 2.8871548938555286e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3365054194768710e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57092, + "real_time": 1.2260314938719997e+04, + "cpu_time": 2.8873243063826751e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3363441381311307e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2259231156987140e+04, + "cpu_time": 2.8884053997057381e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3364623298004982e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2259492711462217e+04, + "cpu_time": 2.8873243063826747e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3364337648882899e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6952749399898996e+00, + "cpu_time": 2.7214181974058071e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9386144054639083e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1985676797143435e-04, + "cpu_time": 9.4218706199727291e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1988007742070612e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57059, + "real_time": 1.2270824070995655e+04, + "cpu_time": 2.9219640477400560e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6703992992168460e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57059, + "real_time": 1.2269717017990344e+04, + "cpu_time": 2.9266209519970598e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6706402398648858e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57059, + "real_time": 1.2262965151123200e+04, + "cpu_time": 2.9195636919679571e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6721106678671989e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57059, + "real_time": 1.2271878894085194e+04, + "cpu_time": 2.9219176045847249e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6701697664074516e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57059, + "real_time": 1.2270538420832327e+04, + "cpu_time": 2.9293355719518288e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6704614643778038e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2269184711005346e+04, + "cpu_time": 2.9238803736483253e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6707562875468373e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2270538420832327e+04, + "cpu_time": 2.9219640477400557e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6704614643778038e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5617737699640939e+00, + "cpu_time": 3.9820995193408926e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.7559356779179128e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9030240018876043e-04, + "cpu_time": 1.3619228595088363e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9040222479610641e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57040, + "real_time": 1.2269548862944950e+04, + "cpu_time": 2.9962817636746113e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3413536823610620e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57040, + "real_time": 1.2268704471768180e+04, + "cpu_time": 2.9999416619915879e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3417212999796772e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57040, + "real_time": 1.2270383841569075e+04, + "cpu_time": 2.9890998352033723e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3409902123827591e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57040, + "real_time": 1.2271850895125537e+04, + "cpu_time": 2.9900360729312779e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3403517171180220e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57040, + "real_time": 1.2271198935568042e+04, + "cpu_time": 2.9889967671809249e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3406354459827108e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2270337401395158e+04, + "cpu_time": 2.9928712201963550e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3410104715648470e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2270383841569075e+04, + "cpu_time": 2.9900360729312779e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3409902123827591e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2573099426227816e+00, + "cpu_time": 4.9723169886469833e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4728482149608352e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0246743031530849e-04, + "cpu_time": 1.6613868833022362e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0246840451067982e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57017, + "real_time": 1.2283274306853786e+04, + "cpu_time": 3.1169210796779997e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0670770409064692e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57017, + "real_time": 1.2265457319427969e+04, + "cpu_time": 3.1219434589683715e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0686270930345781e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57017, + "real_time": 1.2275615143631721e+04, + "cpu_time": 3.1178717540382651e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0677428256456610e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57017, + "real_time": 1.2275160683021224e+04, + "cpu_time": 3.1186930336566289e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0677823564566156e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57017, + "real_time": 1.2277971962727957e+04, + "cpu_time": 3.1194118683901252e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0675378669856323e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2275495883132533e+04, + "cpu_time": 3.1189682389462781e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0677534366057913e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2275615143631720e+04, + "cpu_time": 3.1186930336566285e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0677428256456610e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4717499414752684e+00, + "cpu_time": 1.9050996974575177e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6307324319026209e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.2720883971522046e-04, + "cpu_time": 6.1081086805203969e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2734388285387052e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55980, + "real_time": 1.2502224271107592e+04, + "cpu_time": 3.2938395105394789e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0967788956226765e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55980, + "real_time": 1.2503779237529803e+04, + "cpu_time": 3.2993425616291497e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0965181407968311e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55980, + "real_time": 1.2494847404097687e+04, + "cpu_time": 3.2943666881029014e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0980168186290123e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55980, + "real_time": 1.2497971250537626e+04, + "cpu_time": 3.2938716720257202e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0974924229300278e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55980, + "real_time": 1.2498137452993402e+04, + "cpu_time": 3.2958429653447783e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0974645301025589e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2499391923253224e+04, + "cpu_time": 3.2954526795284051e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0972541616162216e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2498137452993404e+04, + "cpu_time": 3.2943666881029014e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0974645301025589e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5887108353647781e+00, + "cpu_time": 2.3218948111325087e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0213650426327754e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8711083366291808e-04, + "cpu_time": 7.0457537611032630e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8710707327873360e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55774, + "real_time": 1.2547736020311517e+04, + "cpu_time": 3.3048482052569365e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 4.1783473859452759e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55774, + "real_time": 1.2548934697670797e+04, + "cpu_time": 3.3088739358841027e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 4.1779482691651337e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55774, + "real_time": 1.2548640023849746e+04, + "cpu_time": 3.3068405565317335e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 4.1780463779624443e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55774, + "real_time": 1.2545332714741106e+04, + "cpu_time": 3.3037689048660708e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 4.1791478306824608e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55774, + "real_time": 1.2541593949525428e+04, + "cpu_time": 3.3054768099831483e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 4.1803936733244263e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2546447481219719e+04, + "cpu_time": 3.3059616825043988e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 4.1787767074159485e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2547736020311517e+04, + "cpu_time": 3.3054768099831483e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 4.1783473859452759e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0599548991389200e+00, + "cpu_time": 1.9708403038252701e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0193526365077004e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4389014529565005e-04, + "cpu_time": 5.9614735229849354e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4393565578622234e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55465, + "real_time": 1.2622851021522703e+04, + "cpu_time": 3.3160449526728502e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 8.3069664548216263e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55465, + "real_time": 1.2610858879747389e+04, + "cpu_time": 3.3211697899576422e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 8.3148658628158737e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55465, + "real_time": 1.2623699276428095e+04, + "cpu_time": 3.3150592950509315e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 8.3064082646358566e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55465, + "real_time": 1.2624760484783088e+04, + "cpu_time": 3.3155043757324507e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 8.3057100470450317e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55465, + "real_time": 1.2629373329553900e+04, + "cpu_time": 3.3147827639051509e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 8.3026764087037888e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2622308598407035e+04, + "cpu_time": 3.3165122354638050e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 8.3073254076044373e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2623699276428093e+04, + "cpu_time": 3.3155043757324507e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 8.3064082646358566e+10, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8783435261058274e+00, + "cpu_time": 2.6471237228846189e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5291015521495208e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.4493545871425535e-04, + "cpu_time": 7.9816491993566429e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4519370915741782e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 54855, + "real_time": 1.2788774196077204e+04, + "cpu_time": 3.3554695579254527e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.6398381641950290e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 54855, + "real_time": 1.2789838389519067e+04, + "cpu_time": 3.3587504712423783e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.6397017195452295e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 54855, + "real_time": 1.2759325876815481e+04, + "cpu_time": 3.3483170868653629e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.6436228843489767e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 54855, + "real_time": 1.2757952457744912e+04, + "cpu_time": 3.3470453176556257e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.6437998236361914e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 54855, + "real_time": 1.2784425712863116e+04, + "cpu_time": 3.3550930051955242e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.6403959372926230e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2776063326603955e+04, + "cpu_time": 3.3529350877768687e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.6414717058036102e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2784425712863116e+04, + "cpu_time": 3.3550930051955234e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 1.6403959372926230e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6042084194528442e+01, + "cpu_time": 5.0228019038168142e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0619341163463086e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2556359329499846e-03, + "cpu_time": 1.4980313582948406e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2561496546398610e-03, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52432, + "real_time": 1.3379048322106972e+04, + "cpu_time": 3.4667116322093199e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1349793341201331e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52432, + "real_time": 1.3352963343092690e+04, + "cpu_time": 3.4651827509917683e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1411035080611212e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52432, + "real_time": 1.3376617632485670e+04, + "cpu_time": 3.4658480393652797e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1355489969407208e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52432, + "real_time": 1.3347282063899487e+04, + "cpu_time": 3.4623636290814677e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1424405207891510e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52432, + "real_time": 1.3374914130522302e+04, + "cpu_time": 3.4635972135337135e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1359483575512189e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3366165098421425e+04, + "cpu_time": 3.4647406530363100e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1380041434924689e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3374914130522302e+04, + "cpu_time": 3.4651827509917683e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1359483575512189e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4854576905107729e+01, + "cpu_time": 1.7510147611284069e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4889459878190237e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1113566827677512e-03, + "cpu_time": 5.0538119197865872e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1118360041220249e-03, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 48505, + "real_time": 1.4472525933675797e+04, + "cpu_time": 3.7737375157200426e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.7962293786468433e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 48505, + "real_time": 1.4416237432042903e+04, + "cpu_time": 3.7733887949695876e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.8188608779116541e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 48505, + "real_time": 1.4436217354249269e+04, + "cpu_time": 3.7691224265539531e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.8108074948946594e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 48505, + "real_time": 1.4405755091836912e+04, + "cpu_time": 3.7691980063910960e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.8230949690054382e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 48505, + "real_time": 1.4464446583381135e+04, + "cpu_time": 3.7744896732295689e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.7994669561973120e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4439036479037202e+04, + "cpu_time": 3.7719872833728499e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.8096919353311816e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4436217354249267e+04, + "cpu_time": 3.7733887949695883e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.8108074948946594e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9165954969134852e+01, + "cpu_time": 2.6113681129505814e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1733988277707369e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0199377577221579e-03, + "cpu_time": 6.9230565131055751e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0197264172215480e-03, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33722, + "real_time": 2.0767915156319945e+04, + "cpu_time": 4.9478985499080532e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 8.0784305375469897e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33722, + "real_time": 2.0762130330183929e+04, + "cpu_time": 4.9471432951782197e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 8.0806813815292004e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33722, + "real_time": 2.0768493532918430e+04, + "cpu_time": 4.9477792331415818e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 8.0782055633490295e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33722, + "real_time": 2.0764284264084552e+04, + "cpu_time": 4.9427558181602210e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 8.0798431511646753e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33722, + "real_time": 2.0766372425028552e+04, + "cpu_time": 4.9462770713480895e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 8.0790306831728369e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0765839141707082e+04, + "cpu_time": 4.9463707935472332e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 8.0792382633525464e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0766372425028556e+04, + "cpu_time": 4.9471432951782197e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 8.0790306831728369e+11, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6368991176249361e+00, + "cpu_time": 2.1208094710318925e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0259701812246358e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2698254569105586e-04, + "cpu_time": 4.2876071357177370e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2698847933207270e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 21432, + "real_time": 3.2660908862904500e+04, + "cpu_time": 7.3497603956699895e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.0273575711210642e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 21432, + "real_time": 3.2659191073034126e+04, + "cpu_time": 7.3513924318775971e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.0274116075001333e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 21432, + "real_time": 3.2660043307253120e+04, + "cpu_time": 7.3457112821948671e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.0273847981257346e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 21432, + "real_time": 3.2659870793524053e+04, + "cpu_time": 7.3455791713325161e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.0273902249072377e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 21432, + "real_time": 3.2648911710485132e+04, + "cpu_time": 7.3474539520344144e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.0277350834093518e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2657785149440188e+04, + "cpu_time": 7.3479794466218766e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.0274558570127043e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2659870793524053e+04, + "cpu_time": 7.3474539520344144e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.0273902249072377e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9980324374299716e+00, + "cpu_time": 2.5496217252552537e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5727559856244606e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5304260269210715e-04, + "cpu_time": 3.4698269691369431e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5307285221938382e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 11732, + "real_time": 5.9709155668970896e+04, + "cpu_time": 1.2458988041254651e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.1239292073070549e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 11732, + "real_time": 5.9742334951388075e+04, + "cpu_time": 1.2465228187862219e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.1233050073219604e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 11732, + "real_time": 5.9725715252046284e+04, + "cpu_time": 1.2458687700306796e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.1236175861066938e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 11732, + "real_time": 5.9720244506525865e+04, + "cpu_time": 1.2466310756904184e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.1237205164601555e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 11732, + "real_time": 5.9707858378504003e+04, + "cpu_time": 1.2460508446982823e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.1239536272525310e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.9721061751487025e+04, + "cpu_time": 1.2461944626662135e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.1237051888896794e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.9720244506525851e+04, + "cpu_time": 1.2460508446982822e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.1237205164601555e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4061968640061712e+01, + "cpu_time": 3.5796803605136347e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6455639176895711e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3546079436056870e-04, + "cpu_time": 2.8724893808747672e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3543220622693957e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 6385, + "real_time": 1.0961413240540093e+05, + "cpu_time": 2.2183920438528009e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.2244564186632817e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 6385, + "real_time": 1.0959466069275090e+05, + "cpu_time": 2.2183887627251368e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.2246739681623721e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 6385, + "real_time": 1.0958491457430588e+05, + "cpu_time": 2.2177868833202700e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.2247828865987886e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 6385, + "real_time": 1.0962656297729722e+05, + "cpu_time": 2.2180983680500992e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.2243175773721504e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 6385, + "real_time": 1.0960308533206001e+05, + "cpu_time": 2.2181986499608343e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.2245798336184241e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0960467119636301e+05, + "cpu_time": 2.2181729415818281e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.2245621368830034e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0960308533206003e+05, + "cpu_time": 2.2181986499608340e+05, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.2245798336184241e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6292742282711835e+01, + "cpu_time": 2.4992383056286158e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8202715708139816e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4865007216273164e-04, + "cpu_time": 1.1267103023294267e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4864672979741927e-04, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3274, + "real_time": 2.1365967180106134e+05, + "cpu_time": 4.2165989156994253e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2563693173222715e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3274, + "real_time": 2.1364706544667509e+05, + "cpu_time": 4.2167810843006015e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2564434500365264e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3274, + "real_time": 2.1363050714057655e+05, + "cpu_time": 4.2180122877214482e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2565408358243508e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3274, + "real_time": 2.1366651773491304e+05, + "cpu_time": 4.2174246273671323e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2563290629046357e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3274, + "real_time": 2.1367370671883228e+05, + "cpu_time": 4.2177047495418414e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2562867941127979e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1365549376841175e+05, + "cpu_time": 4.2173043329260900e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2563938920401165e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1365967180106131e+05, + "cpu_time": 4.2174246273671323e+05, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.2563693173222715e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7073835874126111e+01, + "cpu_time": 6.0154110441547957e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0040564159152652e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.9912926988121385e-05, + "cpu_time": 1.4263639920861787e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.9915735206646956e-05, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1591, + "real_time": 4.3938768061219191e+05, + "cpu_time": 8.3659965870522987e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.2218615488991094e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1591, + "real_time": 4.3965748707792535e+05, + "cpu_time": 8.3683458705216355e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.2211117239653523e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1591, + "real_time": 4.3992121742849809e+05, + "cpu_time": 8.3714079321181134e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.2203796742021416e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1591, + "real_time": 4.3989123848921340e+05, + "cpu_time": 8.3709446448774368e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.2204628440517681e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1591, + "real_time": 4.4068690507346683e+05, + "cpu_time": 8.3793852357007575e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.2182592807256172e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3990890573625907e+05, + "cpu_time": 8.3712160540540493e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.2204150143687976e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3989123848921340e+05, + "cpu_time": 8.3709446448774380e+05, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.2204628440517681e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8493134812537539e+02, + "cpu_time": 5.0578354627108746e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3443019152510383e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1023449214190464e-03, + "cpu_time": 6.0419363567392858e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1015121081137430e-03, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 718, + "real_time": 9.6656988675817638e+05, + "cpu_time": 1.7434463676880449e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1108786221359250e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 718, + "real_time": 9.6408181269601267e+05, + "cpu_time": 1.7408987618384555e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1137455450978044e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 718, + "real_time": 9.6709043430268229e+05, + "cpu_time": 1.7440616253481682e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1102806789462439e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 718, + "real_time": 9.6376096696170012e+05, + "cpu_time": 1.7406720947075102e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1141163222090427e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 718, + "real_time": 9.6939290407720779e+05, + "cpu_time": 1.7464233440111112e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1076435772161184e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6617920095915569e+05, + "cpu_time": 1.7431004387186584e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1113329491210271e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6656988675817638e+05, + "cpu_time": 1.7434463676880449e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1108786221359250e+12, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3215168814992817e+03, + "cpu_time": 2.3890103543675482e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6689387251170511e+09, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4027808497581403e-03, + "cpu_time": 1.3705523223456330e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4015653699712243e-03, + "own_gpu": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57071, + "real_time": 1.2277640844908839e+04, + "cpu_time": 3.1259668483117526e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3361458049968004e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57071, + "real_time": 1.2272897514279286e+04, + "cpu_time": 3.1260137039827576e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3374351861362654e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57071, + "real_time": 1.2287223178871951e+04, + "cpu_time": 3.1233036393264665e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3335440728732979e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57071, + "real_time": 1.2282533430852396e+04, + "cpu_time": 3.1248157540607521e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3348168951132596e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57071, + "real_time": 1.2272917350924306e+04, + "cpu_time": 3.1256994883565676e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3374297918591613e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2278642463967357e+04, + "cpu_time": 3.1251598868076595e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3358743501957572e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time_median", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2277640844908841e+04, + "cpu_time": 3.1256994883565680e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3361458049968004e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2360423866493466e+00, + "cpu_time": 1.1440544700576760e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6939376434804205e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.0787718633794432e-04, + "cpu_time": 3.6607870044893089e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0779419895746855e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56885, + "real_time": 1.2294333634735733e+04, + "cpu_time": 3.1233921051243571e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6632322201300728e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56885, + "real_time": 1.2308327133311632e+04, + "cpu_time": 3.1220567179397211e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6556567040121329e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56885, + "real_time": 1.2287549708702905e+04, + "cpu_time": 3.1235735483870911e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6669109742830586e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56885, + "real_time": 1.2279012985039040e+04, + "cpu_time": 3.1256319398787084e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6715460029085994e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56885, + "real_time": 1.2274431025620728e+04, + "cpu_time": 3.1261132688757873e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6740364444597340e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2288730897482008e+04, + "cpu_time": 3.1241535160411331e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6662764691587210e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time_median", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2287549708702905e+04, + "cpu_time": 3.1235735483870911e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6669109742830586e+08, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3376495130333689e+01, + "cpu_time": 1.6836745201320547e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.2531831494868512e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0885172148309120e-03, + "cpu_time": 5.3892182682033327e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0880411550651149e-03, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57050, + "real_time": 1.2264766954727169e+04, + "cpu_time": 3.1442060911481138e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3358590554943378e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57050, + "real_time": 1.2269391099723452e+04, + "cpu_time": 3.1461222716914748e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3353555907407084e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57050, + "real_time": 1.2269643821661888e+04, + "cpu_time": 3.1412541051708897e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3353280859770577e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57050, + "real_time": 1.2272824415041428e+04, + "cpu_time": 3.1484896739702130e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3349820258098016e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57050, + "real_time": 1.2269082896647336e+04, + "cpu_time": 3.1457877265556850e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3353891352773492e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2269141837560255e+04, + "cpu_time": 3.1451719737072755e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3353827786598511e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time_median", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2269391099723452e+04, + "cpu_time": 3.1457877265556854e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3353555907407084e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8730893750399202e+00, + "cpu_time": 2.6726380153239852e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1273525007584289e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3417199125079478e-04, + "cpu_time": 8.4975894407888131e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3419146560336380e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56968, + "real_time": 1.2285648307072537e+04, + "cpu_time": 3.2065590858025411e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6671771143844557e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56968, + "real_time": 1.2286419824413022e+04, + "cpu_time": 3.2038523170902816e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6670096308194051e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56968, + "real_time": 1.2285044529815725e+04, + "cpu_time": 3.2044383232692071e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6673081990441527e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56968, + "real_time": 1.2290796804225180e+04, + "cpu_time": 3.2081473581659942e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6660598594173665e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56968, + "real_time": 1.2290961474093709e+04, + "cpu_time": 3.2047750035107641e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6660241405090075e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2287774187924035e+04, + "cpu_time": 3.2055544175677576e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6667157888348780e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time_median", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2286419824413024e+04, + "cpu_time": 3.2047750035107641e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6670096308194051e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8766171835801675e+00, + "cpu_time": 1.7665643207570408e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2424655129203561e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3410400773861870e-04, + "cpu_time": 5.5109478443901658e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3408814464055685e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56912, + "real_time": 1.2302161673096774e+04, + "cpu_time": 3.2745299638037795e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3271938494613266e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56912, + "real_time": 1.2300826157409980e+04, + "cpu_time": 3.2722652252600263e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3277722293897562e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56912, + "real_time": 1.2301928927486528e+04, + "cpu_time": 3.2755723327241994e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3272946369874697e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56912, + "real_time": 1.2301014995418403e+04, + "cpu_time": 3.2699490318386303e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3276904405375757e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56912, + "real_time": 1.2302790802968519e+04, + "cpu_time": 3.2722134189626180e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3269214318581228e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2301744511276040e+04, + "cpu_time": 3.2729059945178509e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3273745176468506e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time_median", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2301928927486528e+04, + "cpu_time": 3.2722652252600259e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3272946369874697e+09, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.1827970079837475e-01, + "cpu_time": 2.2011650374966759e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5436229144760873e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.6517370771952068e-05, + "cpu_time": 6.7254147879091211e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6517247900215912e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56875, + "real_time": 1.2306820342808471e+04, + "cpu_time": 3.4105317679120497e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0650354547232204e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56875, + "real_time": 1.2305855699178410e+04, + "cpu_time": 3.4084221415384622e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0651189417794889e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56875, + "real_time": 1.2306690897493863e+04, + "cpu_time": 3.4071219604395359e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0650466570724672e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56875, + "real_time": 1.2303706587888404e+04, + "cpu_time": 3.4047068641758509e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0653049880840416e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56875, + "real_time": 1.2304086720885627e+04, + "cpu_time": 3.4094239665934198e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0652720756389927e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2305432049650957e+04, + "cpu_time": 3.4080413401318641e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0651556234596420e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time_median", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2305855699178410e+04, + "cpu_time": 3.4084221415384622e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0651189417794889e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4559071359044458e+00, + "cpu_time": 2.2484931109197458e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2602643247827021e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1831418271459577e-04, + "cpu_time": 6.5976110220328113e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1831738921767545e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 55847, + "real_time": 1.2536411318272429e+04, + "cpu_time": 3.5717081759091678e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0910609371751579e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 55847, + "real_time": 1.2537324150388355e+04, + "cpu_time": 3.5722505649363477e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0909086887721558e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 55847, + "real_time": 1.2532403422286310e+04, + "cpu_time": 3.5649655003850195e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0917296640310081e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 55847, + "real_time": 1.2531852849090365e+04, + "cpu_time": 3.5715396171683518e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0918215618772446e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 55847, + "real_time": 1.2535225937984915e+04, + "cpu_time": 3.5650308539402431e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0912586761251518e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2534643535604475e+04, + "cpu_time": 3.5690989424678257e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0913559055961441e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time_median", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2535225937984917e+04, + "cpu_time": 3.5715396171683518e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.0912586761251518e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4215688324781577e+00, + "cpu_time": 3.7527457965090896e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0404020748831420e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9319008359509597e-04, + "cpu_time": 1.0514546828209483e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9319533629219457e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 44386, + "real_time": 1.5772299782816379e+04, + "cpu_time": 3.8948603681340777e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.3241062319345581e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 44386, + "real_time": 1.5772660998966487e+04, + "cpu_time": 3.8997039269138957e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.3240301052203827e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 44386, + "real_time": 1.5771001896235761e+04, + "cpu_time": 3.8961166110935534e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.3243797917818882e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 44386, + "real_time": 1.5768728159214950e+04, + "cpu_time": 3.8952763709278101e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.3248591434028614e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 44386, + "real_time": 1.5768194048257274e+04, + "cpu_time": 3.8946133871040380e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.3249717652856079e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5770576977098168e+04, + "cpu_time": 3.8961141328346755e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.3244694075250599e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time_median", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5771001896235761e+04, + "cpu_time": 3.8952763709278093e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.3243797917818882e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0364331762274590e+00, + "cpu_time": 2.0864225569340825e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2929342547772611e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2912864121488650e-04, + "cpu_time": 5.3551371592291507e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2913141101735047e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 36315, + "real_time": 1.9276615690533417e+04, + "cpu_time": 4.2949060415806278e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4396270425982857e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 36315, + "real_time": 1.9284122964143848e+04, + "cpu_time": 4.2897540410299087e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4375094057929497e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 36315, + "real_time": 1.9287484466328995e+04, + "cpu_time": 4.2904251604020399e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4365617342714905e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 36315, + "real_time": 1.9284951663558837e+04, + "cpu_time": 4.2932462839047512e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4372757489530380e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 36315, + "real_time": 1.9287309239470826e+04, + "cpu_time": 4.2881111937216490e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4366111259009872e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9284096804807185e+04, + "cpu_time": 4.2912885441277955e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4375170115033508e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time_median", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9284951663558833e+04, + "cpu_time": 4.2904251604020399e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 5.4372757489530380e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4298568109743064e+00, + "cpu_time": 2.7438440038192532e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2493678369120922e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2971554518799248e-04, + "cpu_time": 6.3939862714986449e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2976807875156792e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 26681, + "real_time": 2.6190905297728208e+04, + "cpu_time": 5.0755335931936082e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 8.0071764460234467e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 26681, + "real_time": 2.6296263931496134e+04, + "cpu_time": 5.0887085229189055e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 7.9750948859626907e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 26681, + "real_time": 2.6152483487309735e+04, + "cpu_time": 5.0699319215922304e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 8.0189401554067520e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 26681, + "real_time": 2.6223216046795089e+04, + "cpu_time": 5.0808688879727531e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 7.9973104605386749e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 26681, + "real_time": 2.6292299064456070e+04, + "cpu_time": 5.0913097072823271e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 7.9762975267350800e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6231033565557049e+04, + "cpu_time": 5.0812705265919649e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 7.9949638949333298e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time_median", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6223216046795085e+04, + "cpu_time": 5.0808688879727531e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 7.9973104605386749e+10, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2948328545389053e+01, + "cpu_time": 8.9127243315764417e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9188047449560010e+08, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3997654681836124e-03, + "cpu_time": 1.7540346031436854e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4000167732739987e-03, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 19019, + "real_time": 3.6769659266737552e+04, + "cpu_time": 6.2624868237026545e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1406969995488202e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 19019, + "real_time": 3.6809508012879167e+04, + "cpu_time": 6.2625958567748457e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1394621190080745e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 19019, + "real_time": 3.6775865334360671e+04, + "cpu_time": 6.2571832693622106e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1405045025768979e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 19019, + "real_time": 3.6808070451178763e+04, + "cpu_time": 6.2648605815236187e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1395066213979384e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 19019, + "real_time": 3.6779079752126709e+04, + "cpu_time": 6.2637193175245331e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1404048247720143e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6788436563456577e+04, + "cpu_time": 6.2621691697775721e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1401150134607491e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time_median", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6779079752126716e+04, + "cpu_time": 6.2625958567748450e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 1.1404048247720143e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8892247869440219e+01, + "cpu_time": 2.9491025746717952e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8540845616512239e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.1353766656685390e-04, + "cpu_time": 4.7093946118618584e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1346438670968027e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10852, + "real_time": 6.4473072750580308e+04, + "cpu_time": 9.2191917250274739e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.3011025598938118e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10852, + "real_time": 6.4500143772211566e+04, + "cpu_time": 9.2154519351273309e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.3005564808700540e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10852, + "real_time": 6.4462797381351353e+04, + "cpu_time": 9.2072973092519911e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.3013099556282623e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10852, + "real_time": 6.4487198115140185e+04, + "cpu_time": 9.2121718208626771e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.3008175645997771e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10852, + "real_time": 6.4465273952521304e+04, + "cpu_time": 9.2083592056765076e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.3012599630272591e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4477697194360953e+04, + "cpu_time": 9.2124943991891967e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.3010093048038330e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time_median", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4473072750580301e+04, + "cpu_time": 9.2121718208626771e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 1.3011025598938118e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5746771053756024e+01, + "cpu_time": 4.9450687096751139e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1769906909718197e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4422043185396504e-04, + "cpu_time": 5.3677847664258393e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4419430969795011e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5811, + "real_time": 1.2049189893751353e+05, + "cpu_time": 1.5421081758733073e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.3923936918531409e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5811, + "real_time": 1.2048655703611216e+05, + "cpu_time": 1.5417860454311068e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.3924554251285925e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5811, + "real_time": 1.2052290124824372e+05, + "cpu_time": 1.5413361400792323e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.3920355240572571e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5811, + "real_time": 1.2047796726793438e+05, + "cpu_time": 1.5414054672173219e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.3925547036072308e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5811, + "real_time": 1.2051687056620503e+05, + "cpu_time": 1.5416694785751528e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.3921051817208914e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2049923901120175e+05, + "cpu_time": 1.5416610614352243e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.3923089052734225e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time_median", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2049189893751351e+05, + "cpu_time": 1.5416694785751525e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.3923936918531409e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9608476618729636e+01, + "cpu_time": 3.1084781678178881e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2655806769797452e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6272697470651087e-04, + "cpu_time": 2.0163174938880667e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6272112233131403e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3052, + "real_time": 2.2932132817998584e+05, + "cpu_time": 2.7450098328964820e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.4632058983045993e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3052, + "real_time": 2.2936111668267913e+05, + "cpu_time": 2.7457633453473327e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.4629520681320419e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3052, + "real_time": 2.2931189210689504e+05, + "cpu_time": 2.7450859600263159e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.4632661085173206e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3052, + "real_time": 2.2935257139468478e+05, + "cpu_time": 2.7450978112712293e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.4630065752459933e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3052, + "real_time": 2.2931063191198985e+05, + "cpu_time": 2.7449119102228485e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.4632741500131705e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2933150805524690e+05, + "cpu_time": 2.7451737719528412e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.4631409600426251e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time_median", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2932132817998584e+05, + "cpu_time": 2.7450859600263153e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 1.4632058983045993e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3688413539330490e+01, + "cpu_time": 3.3783328897811558e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5112752490486966e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0329332301614592e-04, + "cpu_time": 1.2306444583935763e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0328979164144713e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1560, + "real_time": 4.4858054420612269e+05, + "cpu_time": 5.1759649038460176e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.4960270762247656e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1560, + "real_time": 4.4854290330803423e+05, + "cpu_time": 5.1759570576924336e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.4961526200741910e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1560, + "real_time": 4.4858454594549799e+05, + "cpu_time": 5.1767579166666936e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.4960137304452209e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1560, + "real_time": 4.4857900718806119e+05, + "cpu_time": 5.1760489615384716e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.4960322022351224e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1560, + "real_time": 4.4857203105586965e+05, + "cpu_time": 5.1759048461537284e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.4960554683277075e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4857180634071713e+05, + "cpu_time": 5.1761267371794692e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.4960562194614014e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time_median", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4857900718806114e+05, + "cpu_time": 5.1759649038460181e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 1.4960322022351224e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6777485026738752e+01, + "cpu_time": 3.5659969132278064e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5957828967178492e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7402005185308602e-05, + "cpu_time": 6.8893153013694535e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7403560266822059e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 788, + "real_time": 8.8849936202428490e+05, + "cpu_time": 1.0049560126904177e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.5106114166948779e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 788, + "real_time": 8.8865387728875072e+05, + "cpu_time": 1.0051605723350114e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.5103487581632254e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 788, + "real_time": 8.8853794616897625e+05, + "cpu_time": 1.0050154403553004e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.5105458194407306e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 788, + "real_time": 8.8858789070580283e+05, + "cpu_time": 1.0051104530456451e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.5104609167405066e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 788, + "real_time": 8.8837205019878782e+05, + "cpu_time": 1.0048128654822101e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.5108279011025458e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8853022527732048e+05, + "cpu_time": 1.0050110687817171e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.5105589624283774e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time_median", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8853794616897614e+05, + "cpu_time": 1.0050154403553003e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.5105458194407306e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0564874510727475e+02, + "cpu_time": 1.3653147594278420e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7961833648079474e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1890281512291893e-04, + "cpu_time": 1.3585071864759539e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1890852389637274e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 396, + "real_time": 1.7686724883852282e+06, + "cpu_time": 1.9812524823232947e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.5177227992338910e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 396, + "real_time": 1.7686021720609544e+06, + "cpu_time": 1.9811426767677062e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.5177831410621408e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 396, + "real_time": 1.7685823879237615e+06, + "cpu_time": 1.9812230303030692e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.5178001196491135e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 396, + "real_time": 1.7688535089190635e+06, + "cpu_time": 1.9813682272727403e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.5175674788583221e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 396, + "real_time": 1.7685120930598879e+06, + "cpu_time": 1.9810724722222832e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.5178604492070602e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7686445300697791e+06, + "cpu_time": 1.9812117777778185e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.5177467976021054e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time_median", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7686021720609546e+06, + "cpu_time": 1.9812230303030692e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 1.5177831410621408e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3004733266850400e+02, + "cpu_time": 1.1229527212991857e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1159291063934125e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.3529378265384503e-05, + "cpu_time": 5.6680095176837690e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.3525380396550574e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 198, + "real_time": 3.5298669350926173e+06, + "cpu_time": 3.9451339949492868e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.5209381029710501e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 198, + "real_time": 3.5298903839606228e+06, + "cpu_time": 3.9459596919192625e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.5209279994627420e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 198, + "real_time": 3.5299461492516026e+06, + "cpu_time": 3.9454572828283929e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.5209039721861594e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 198, + "real_time": 3.5300528139082920e+06, + "cpu_time": 3.9458166414143611e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.5208580163014737e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 198, + "real_time": 3.5296156285142507e+06, + "cpu_time": 3.9450992878787224e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.5210463928787320e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5298743821454770e+06, + "cpu_time": 3.9454933797980053e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.5209348967600314e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time_median", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5298903839606224e+06, + "cpu_time": 3.9454572828283934e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 1.5209279994627420e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6144145994309068e+02, + "cpu_time": 3.8978005452123716e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9563230040704692e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.5735752172848056e-05, + "cpu_time": 9.8791207334681281e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5737151661712559e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 99, + "real_time": 7.0598146389003349e+06, + "cpu_time": 7.8621132828281969e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.5209207024835574e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 99, + "real_time": 7.0598340978977652e+06, + "cpu_time": 7.8623120303031690e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.5209165103748434e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 99, + "real_time": 7.0593153512237044e+06, + "cpu_time": 7.8616253939393843e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.5210282733917972e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 99, + "real_time": 7.0600135562320547e+06, + "cpu_time": 7.8624813434347827e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.5208778502304440e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 99, + "real_time": 7.0593719643474827e+06, + "cpu_time": 7.8617077676766468e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.5210160753999155e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0596699217202682e+06, + "cpu_time": 7.8620479636364356e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.5209518823761118e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time_median", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0598146389003349e+06, + "cpu_time": 7.8621132828281969e+06, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.5209207024835574e+11, + "own_gpu": 1.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0841546763575195e+02, + "cpu_time": 3.7286535524646695e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6446313741185069e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.3686952939096998e-05, + "cpu_time": 4.7425983277009342e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3687321414388936e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56920, + "real_time": 1.2292831400357622e+04, + "cpu_time": 3.3593948208011454e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3320232472079945e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56920, + "real_time": 1.2298638666061090e+04, + "cpu_time": 3.3578522189036572e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3304499068691105e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56920, + "real_time": 1.2292122141763828e+04, + "cpu_time": 3.3640695555165403e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3322155058022022e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56920, + "real_time": 1.2297850980710149e+04, + "cpu_time": 3.3649012473647243e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3306632243509859e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56920, + "real_time": 1.2292546158547844e+04, + "cpu_time": 3.3645921626844691e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3321005649848807e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2294797869488106e+04, + "cpu_time": 3.3621620010541068e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3314904898430347e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time_median", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2292831400357622e+04, + "cpu_time": 3.3640695555165395e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3320232472079945e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1689875490224262e+00, + "cpu_time": 3.2893551289839152e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.5861099457204720e+04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5775027639021822e-04, + "cpu_time": 9.7834522189966882e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5772578285597963e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56880, + "real_time": 1.2292663033964811e+04, + "cpu_time": 3.3616116121659070e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6641377684927845e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56880, + "real_time": 1.2296356114191813e+04, + "cpu_time": 3.3598726969057701e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6621362653487408e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56880, + "real_time": 1.2292778448970383e+04, + "cpu_time": 3.3571382260899525e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6640751999285758e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56880, + "real_time": 1.2291695361838667e+04, + "cpu_time": 3.3579836409986769e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6646624072975647e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56880, + "real_time": 1.2292232794298196e+04, + "cpu_time": 3.3592787201125066e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6643710195595181e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2293145150652776e+04, + "cpu_time": 3.3591769792545630e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6638765321254361e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time_median", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2292663033964811e+04, + "cpu_time": 3.3592787201125066e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6641377684927845e+08, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8446866424351249e+00, + "cpu_time": 1.7322227285104393e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9979490696842418e+04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5005815190729857e-04, + "cpu_time": 5.1566879006620186e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5003202747658663e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56923, + "real_time": 1.2290213490692378e+04, + "cpu_time": 3.3944966252656908e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3330931974784594e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56923, + "real_time": 1.2300892077724258e+04, + "cpu_time": 3.3917151836690886e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3319359194826090e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56923, + "real_time": 1.2288445227056443e+04, + "cpu_time": 3.3947453226288388e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3332850248562000e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56923, + "real_time": 1.2290297949809090e+04, + "cpu_time": 3.3925360065351815e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3330840364414842e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56923, + "real_time": 1.2288406840318115e+04, + "cpu_time": 3.3942855770075621e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3332891897950754e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2291651117120058e+04, + "cpu_time": 3.3935557430212721e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3329374736107657e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time_median", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2290213490692378e+04, + "cpu_time": 3.3942855770075621e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3330931974784594e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2463379454568839e+00, + "cpu_time": 1.3472692673624485e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6862496434820723e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.2682125415597580e-04, + "cpu_time": 3.9700814407809870e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2659537720690774e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56904, + "real_time": 1.2294262092359062e+04, + "cpu_time": 3.4505366986503104e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6653083978391399e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56904, + "real_time": 1.2299164623591500e+04, + "cpu_time": 3.4491556973148348e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6642459876621575e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56904, + "real_time": 1.2296246221572083e+04, + "cpu_time": 3.4357238383945252e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6648783221755123e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56904, + "real_time": 1.2303141124065543e+04, + "cpu_time": 3.4349135526501428e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6633848762332897e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56904, + "real_time": 1.2295186545966311e+04, + "cpu_time": 3.4404491705328248e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6651079979547133e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2297600121510899e+04, + "cpu_time": 3.4421557915085279e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6645851163729630e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time_median", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2296246221572084e+04, + "cpu_time": 3.4404491705328255e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6648783221755123e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6038592759963599e+00, + "cpu_time": 7.3479739641195323e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8072010656828864e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9305386745276477e-04, + "cpu_time": 2.1347011608964035e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9299874932537566e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56889, + "real_time": 1.2282118737244182e+04, + "cpu_time": 3.5185216087468230e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3358871870591230e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56889, + "real_time": 1.2292353370939210e+04, + "cpu_time": 3.5178188208616659e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3314445185846987e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56889, + "real_time": 1.2311689291352342e+04, + "cpu_time": 3.5083824289405035e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3230713063910818e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56889, + "real_time": 1.2313950955627355e+04, + "cpu_time": 3.5101654502628422e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3220936347850800e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56889, + "real_time": 1.2303825006589155e+04, + "cpu_time": 3.5097219480039326e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3264736750484533e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2300787472350450e+04, + "cpu_time": 3.5129220513631539e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3277940643736877e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time_median", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2303825006589157e+04, + "cpu_time": 3.5101654502628415e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3264736750484533e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3425808177595377e+01, + "cpu_time": 4.8420315281056155e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8173709517066209e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0914592425707488e-03, + "cpu_time": 1.3783486958461586e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0918911056654150e-03, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 45359, + "real_time": 1.5425051566254304e+04, + "cpu_time": 3.9742926497498171e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4973459853287525e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 45359, + "real_time": 1.5432360785901743e+04, + "cpu_time": 3.9724662999625391e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4933213925209045e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 45359, + "real_time": 1.5433483098373576e+04, + "cpu_time": 3.9753569038118367e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4927037639230471e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 45359, + "real_time": 1.5431807598176016e+04, + "cpu_time": 3.9759457197028481e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4936258546595821e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 45359, + "real_time": 1.5425750854371594e+04, + "cpu_time": 3.9780602702882468e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4969607792449694e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5429690780615445e+04, + "cpu_time": 3.9752243687030583e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4947915551354523e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time_median", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5431807598176016e+04, + "cpu_time": 3.9753569038118367e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.4936258546595821e+09, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9697873844404810e+00, + "cpu_time": 2.0649027819176432e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1857352649211660e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5728236818768814e-04, + "cpu_time": 5.1944307802463251e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5730298980671266e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 36955, + "real_time": 1.8927205759773100e+04, + "cpu_time": 4.5109470382897671e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3850116246802114e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 36955, + "real_time": 1.8934639961343808e+04, + "cpu_time": 4.5079583168719357e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3844678353281738e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 36955, + "real_time": 1.8928166315631352e+04, + "cpu_time": 4.5131611744014284e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3849413388951202e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 36955, + "real_time": 1.8936076384124946e+04, + "cpu_time": 4.5133798511703768e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3843628145679026e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 36955, + "real_time": 1.8929998381036250e+04, + "cpu_time": 4.5142496468677215e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3848073027972965e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8931217360381892e+04, + "cpu_time": 4.5119392055202465e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3847181832537409e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time_median", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8929998381036250e+04, + "cpu_time": 4.5131611744014284e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3848073027972965e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9437302576040105e+00, + "cpu_time": 2.5366650005274629e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8844825915231314e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0831889373671293e-04, + "cpu_time": 5.6221169767179400e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0830827719365392e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27431, + "real_time": 2.5512716152978366e+04, + "cpu_time": 5.1737283000984942e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0550065969310539e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27431, + "real_time": 2.5519459175741729e+04, + "cpu_time": 5.1759783748314672e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0544636012442513e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27431, + "real_time": 2.5519085289900027e+04, + "cpu_time": 5.1780757792280347e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0544937016512238e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27431, + "real_time": 2.5519693624119533e+04, + "cpu_time": 5.1791431446174000e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0544447269714775e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27431, + "real_time": 2.5511040138091495e+04, + "cpu_time": 5.1772469468849617e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0551416059949898e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5516398876166233e+04, + "cpu_time": 5.1768345091320720e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0547100465585995e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time_median", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5519085289900024e+04, + "cpu_time": 5.1772469468849617e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0544937016512238e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1748095957868472e+00, + "cpu_time": 2.0869927323734405e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3619955491225743e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6361280508459040e-04, + "cpu_time": 4.0314070861101131e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6362384341058370e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18102, + "real_time": 3.8658933983339870e+04, + "cpu_time": 6.4984051927961678e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.7123769125446796e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18102, + "real_time": 3.8671114038680309e+04, + "cpu_time": 6.4967595293339626e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.7115226081958092e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18102, + "real_time": 3.8672971138233072e+04, + "cpu_time": 6.5017900121531027e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.7113923992339741e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18102, + "real_time": 3.8675233108531691e+04, + "cpu_time": 6.5000666887636333e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.7112338199939278e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18102, + "real_time": 3.8667897323027602e+04, + "cpu_time": 6.5025142138989257e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.7117481750825623e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8669229918362507e+04, + "cpu_time": 6.4999071273891590e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.7116547830101906e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time_median", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8671114038680309e+04, + "cpu_time": 6.5000666887636318e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.7115226081958092e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3515158886966843e+00, + "cpu_time": 2.3733307803977596e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4545438321605949e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6425245349094986e-04, + "cpu_time": 3.6513302942392409e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6427400198839597e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10845, + "real_time": 6.4631122122671142e+04, + "cpu_time": 9.1186547349011365e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.2448020877922623e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10845, + "real_time": 6.4574275153049100e+04, + "cpu_time": 9.1147263992626613e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.2476585993872757e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10845, + "real_time": 6.4572829296489312e+04, + "cpu_time": 9.1213701152602691e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.2477313180297924e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10845, + "real_time": 6.4614537204920634e+04, + "cpu_time": 9.1211377685572152e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.2456349464347694e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10845, + "real_time": 6.4617680157628165e+04, + "cpu_time": 9.1185547810051852e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.2454770813254421e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4602088786951688e+04, + "cpu_time": 9.1188887597972949e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.2462608065939087e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time_median", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4614537204920627e+04, + "cpu_time": 9.1186547349011365e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.2456349464347694e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6789478071597465e+01, + "cpu_time": 2.6789326236948892e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3462926274476882e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.1468439449295958e-04, + "cpu_time": 2.9377840812200453e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1472103064333450e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 6161, + "real_time": 1.1372123013518544e+05, + "cpu_time": 1.4117586933940114e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.6882330546495552e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 6161, + "real_time": 1.1362994336573756e+05, + "cpu_time": 1.4118877130336355e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.6911960666035973e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 6161, + "real_time": 1.1374693950644032e+05, + "cpu_time": 1.4119160412270151e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.6873994308765724e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 6161, + "real_time": 1.1364973368229141e+05, + "cpu_time": 1.4116518065249579e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.6905533027690193e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 6161, + "real_time": 1.1376961179661989e+05, + "cpu_time": 1.4122804642104005e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.6866645967799759e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1370349169725491e+05, + "cpu_time": 1.4118989436780041e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.6888092903357437e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time_median", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1372123013518544e+05, + "cpu_time": 1.4118877130336355e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.6882330546495552e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.0978506613839670e+01, + "cpu_time": 2.3807734260845073e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9784932415844135e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3629405485805190e-04, + "cpu_time": 1.6862208423235874e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3635010266533392e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3176, + "real_time": 2.2040217107966670e+05, + "cpu_time": 2.5010816152393899e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8060459926085976e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3176, + "real_time": 2.2034609892605850e+05, + "cpu_time": 2.5008138287154411e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8070145289093422e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3176, + "real_time": 2.2037501889654851e+05, + "cpu_time": 2.5009590018890746e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8065149316846550e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3176, + "real_time": 2.2031305018867840e+05, + "cpu_time": 2.5014052707807615e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8075856118445587e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3176, + "real_time": 2.2035371136463949e+05, + "cpu_time": 2.5019492065490381e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8068830100703873e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2035801009111837e+05, + "cpu_time": 2.5012417846347415e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8068088150235085e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time_median", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2035371136463952e+05, + "cpu_time": 2.5010816152393896e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.8068830100703873e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3246202969606685e+01, + "cpu_time": 4.5164700496712477e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7434785849135015e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5087358501676129e-04, + "cpu_time": 1.8056911080792582e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5087383853496812e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1611, + "real_time": 4.3429186430942989e+05, + "cpu_time": 4.6921537678460189e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.8631200302767708e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1611, + "real_time": 4.3441144389147777e+05, + "cpu_time": 4.6929840657974296e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.8620566368392426e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1611, + "real_time": 4.3432502654348721e+05, + "cpu_time": 4.6921515270018799e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.8628250675580551e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1611, + "real_time": 4.3447978275655792e+05, + "cpu_time": 4.6958553693359555e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.8614491780392906e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1611, + "real_time": 4.3435681981357158e+05, + "cpu_time": 4.6930586902544211e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.8625423234291283e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3437298746290489e+05, + "cpu_time": 4.6932406840471411e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.8623986472284973e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time_median", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3435681981357158e+05, + "cpu_time": 4.6929840657974296e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.8625423234291283e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4207041886842305e+01, + "cpu_time": 1.5250576812203775e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5979798337612404e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7083714694201494e-04, + "cpu_time": 3.2494768197254865e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7082596687671500e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 810, + "real_time": 8.6374341238600516e+05, + "cpu_time": 9.1021940000001888e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.8847684993983597e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 810, + "real_time": 8.6376455922921502e+05, + "cpu_time": 9.1020751358021214e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.8846733917796394e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 810, + "real_time": 8.6370253695439501e+05, + "cpu_time": 9.1022087283952849e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.8849523492567596e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 810, + "real_time": 8.6375388530956826e+05, + "cpu_time": 9.1023567901235190e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.8847213969954109e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 810, + "real_time": 8.6374106906055496e+05, + "cpu_time": 9.1045422345681244e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.8847790387569931e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6374109258794761e+05, + "cpu_time": 9.1026753777778486e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.8847789352374329e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time_median", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6374341238600505e+05, + "cpu_time": 9.1022087283952837e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.8847684993983597e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3485397781227213e+01, + "cpu_time": 1.0483831532342508e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0563073104357463e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7190321246451392e-05, + "cpu_time": 1.1517307931178612e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7190924581431455e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 406, + "real_time": 1.7255519994470321e+06, + "cpu_time": 1.7959539285714119e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.8891244089720627e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 406, + "real_time": 1.7254077694654060e+06, + "cpu_time": 1.7956824088669752e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.8894495079729912e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 406, + "real_time": 1.7254050156408965e+06, + "cpu_time": 1.7957112561576222e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.8894557157104706e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 406, + "real_time": 1.7253817563757226e+06, + "cpu_time": 1.7956332290641242e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.8895081480962547e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 406, + "real_time": 1.7254751616070424e+06, + "cpu_time": 1.7958247438423259e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.8892975971614304e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7254443405072200e+06, + "cpu_time": 1.7957611133004918e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.8893670755826416e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time_median", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7254077694654062e+06, + "cpu_time": 1.7957112561576224e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.8894495079729912e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9536047182977867e+01, + "cpu_time": 1.2870613212338361e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5673862112957356e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.0300371069945208e-05, + "cpu_time": 7.1672190231823282e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0299261572294134e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 203, + "real_time": 3.4507808592078723e+06, + "cpu_time": 3.5766368029556745e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.8894885962364395e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 203, + "real_time": 3.4508391930331737e+06, + "cpu_time": 3.5766266847291114e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.8894228473749031e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 203, + "real_time": 3.4507918389504915e+06, + "cpu_time": 3.5764865172413448e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.8894762206468063e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 203, + "real_time": 3.4507123037732295e+06, + "cpu_time": 3.5764318817734816e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.8895658688566345e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 203, + "real_time": 3.4504293124627303e+06, + "cpu_time": 3.5759988522166172e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.8898848765054871e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4507107014854997e+06, + "cpu_time": 3.5764361477832459e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.8895676819240540e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time_median", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4507808592078723e+06, + "cpu_time": 3.5764865172413453e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.8894885962364395e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6370539088065300e+02, + "cpu_time": 2.5997572830923252e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8453486143490612e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.7441065056592351e-05, + "cpu_time": 7.2691281926107133e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7443540394602974e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 101, + "real_time": 6.9035427451207498e+06, + "cpu_time": 7.1380244950498110e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.8883724764320961e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 101, + "real_time": 6.9035934007698949e+06, + "cpu_time": 7.1382870693072947e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.8883439451990875e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 101, + "real_time": 6.9037185345619619e+06, + "cpu_time": 7.1383124554458987e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.8882734667720940e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 101, + "real_time": 6.9037137580758864e+06, + "cpu_time": 7.1380272772278730e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.8882761569595383e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 101, + "real_time": 6.9036852697482202e+06, + "cpu_time": 7.1380707227722229e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.8882922020833946e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9036507416553423e+06, + "cpu_time": 7.1381444039606210e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.8883116494892426e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time_median", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9036852697482212e+06, + "cpu_time": 7.1380707227722229e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.8882922020833946e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8597678121107862e+01, + "cpu_time": 1.4328330774299025e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4268388721524528e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1384944149457638e-05, + "cpu_time": 2.0072906855665330e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1384989865032936e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.3815276288226539e+07, + "cpu_time": 1.4239926000000507e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.8860671390084656e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.3815957103289809e+07, + "cpu_time": 1.4241023941176359e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.8858756435496033e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.3815558679840144e+07, + "cpu_time": 1.4240406313725615e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.8859877073477280e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3815762602961531e+07, + "cpu_time": 1.4240802549020195e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.8859303494757286e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.3815696771238364e+07, + "cpu_time": 1.4240181980391761e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.8859488659136070e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3815650289111277e+07, + "cpu_time": 1.4240468156862888e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.8859619410590263e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time_median", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3815696771238366e+07, + "cpu_time": 1.4240406313725615e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.8859488659136070e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5356401123822207e+02, + "cpu_time": 4.4731690068800481e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1321044075363898e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8353389520728317e-05, + "cpu_time": 3.1411671004118644e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8353510702661962e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 25, + "real_time": 2.7648043558001518e+07, + "cpu_time": 2.8452818040000238e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.8836086963891243e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 25, + "real_time": 2.7645643875002861e+07, + "cpu_time": 2.8449220920001607e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.8839457993990700e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 25, + "real_time": 2.7648011744022369e+07, + "cpu_time": 2.8451291960000161e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.8836131651750618e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 25, + "real_time": 2.7646424323320389e+07, + "cpu_time": 2.8449870919998832e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.8838361570478912e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 25, + "real_time": 2.7647467628121376e+07, + "cpu_time": 2.8452427760000773e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.8836895966119263e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7647118225693710e+07, + "cpu_time": 2.8451125920000326e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.8837386829246147e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time_median", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7647467628121376e+07, + "cpu_time": 2.8451291960000165e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.8836895966119263e+10, + "own_gpu": 2.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0526047265614156e+03, + "cpu_time": 1.5644229467206942e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4786744818789563e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.8072855115264158e-05, + "cpu_time": 5.4986328172726195e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8073480287954230e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57223, + "real_time": 1.2218297002052994e+04, + "cpu_time": 3.1227803558009626e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3523493489409900e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57223, + "real_time": 1.2241674056273891e+04, + "cpu_time": 3.1218432920329960e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3459476058348316e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57223, + "real_time": 1.2253973484322350e+04, + "cpu_time": 3.1225519598762545e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3425892468597180e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57223, + "real_time": 1.2297356670181845e+04, + "cpu_time": 3.1206267602187872e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3307971053094870e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57223, + "real_time": 1.2245617605830794e+04, + "cpu_time": 3.1205957691836458e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3448700848291028e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2251383763732378e+04, + "cpu_time": 3.1216796274225293e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3433106783548260e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time_median", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2245617605830794e+04, + "cpu_time": 3.1218432920329960e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3448700848291028e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8907073417635235e+01, + "cpu_time": 1.0347259310600970e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8770659959149768e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3594945660920770e-03, + "cpu_time": 3.3146448532722660e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3560676089459678e-03, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57301, + "real_time": 1.2230425359162347e+04, + "cpu_time": 3.1214217203887565e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6980499528276944e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57301, + "real_time": 1.2258070462540942e+04, + "cpu_time": 3.1207473098200800e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6829441265113282e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57301, + "real_time": 1.2278142670879843e+04, + "cpu_time": 3.1197549955498202e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6720189035016060e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57301, + "real_time": 1.2270865367205246e+04, + "cpu_time": 3.1190566639325567e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6759757807250488e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57301, + "real_time": 1.2234233505966160e+04, + "cpu_time": 3.1194391895430032e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6959650524939549e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2254347473150907e+04, + "cpu_time": 3.1200839758468432e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6849907632119274e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time_median", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2258070462540943e+04, + "cpu_time": 3.1197549955498202e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6829441265113282e+08, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1387800977129206e+01, + "cpu_time": 9.7582267878611919e+00, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1669460505781746e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7453235289752931e-03, + "cpu_time": 3.1275526118532257e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7456210366063301e-03, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57243, + "real_time": 1.2228846770467244e+04, + "cpu_time": 3.1487548521216675e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3397829171895001e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57243, + "real_time": 1.2229510060490220e+04, + "cpu_time": 3.1459743444613272e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3397102515931246e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57243, + "real_time": 1.2223969864640747e+04, + "cpu_time": 3.1492761717590834e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3403174403589315e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57243, + "real_time": 1.2228649486889286e+04, + "cpu_time": 3.1488322816763684e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3398045317731769e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57243, + "real_time": 1.2227821930140206e+04, + "cpu_time": 3.1469686302255403e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3398952073071396e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2227759622525544e+04, + "cpu_time": 3.1479612560487974e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3399020696443746e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time_median", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2228649486889288e+04, + "cpu_time": 3.1487548521216675e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3398045317731769e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2024945622214540e+00, + "cpu_time": 1.4186796129560184e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4139440258630688e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8012249424367952e-04, + "cpu_time": 4.5066616059204353e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8015824294560252e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57181, + "real_time": 1.2241250807470449e+04, + "cpu_time": 3.1840552561165041e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6768506352310600e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57181, + "real_time": 1.2239459883710932e+04, + "cpu_time": 3.1812664433989568e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6772423220741777e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57181, + "real_time": 1.2242787732554811e+04, + "cpu_time": 3.1807464857208233e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6765145909429250e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57181, + "real_time": 1.2241469186674172e+04, + "cpu_time": 3.1782470173659091e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6768028820977316e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57181, + "real_time": 1.2240865417118894e+04, + "cpu_time": 3.1805134695091194e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6769349129656987e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2241166605505852e+04, + "cpu_time": 3.1809657344222622e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6768690686623187e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time_median", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2241250807470449e+04, + "cpu_time": 3.1807464857208233e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6768506352310600e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1969034770715097e+00, + "cpu_time": 2.0784372839232610e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6173865194120642e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.7776912580632990e-05, + "cpu_time": 6.5339819961963638e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.7777905914539967e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57070, + "real_time": 1.2262352202427635e+04, + "cpu_time": 3.2592112142981889e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3444884731842499e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57070, + "real_time": 1.2264403995749364e+04, + "cpu_time": 3.2577696285264454e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3435943583327551e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57070, + "real_time": 1.2262685809307412e+04, + "cpu_time": 3.2578722393551761e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3443430761520443e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57070, + "real_time": 1.2263534835950350e+04, + "cpu_time": 3.2568962186787670e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3439730776384563e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57070, + "real_time": 1.2263602209874012e+04, + "cpu_time": 3.2574574610128213e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3439437188556099e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2263315810661756e+04, + "cpu_time": 3.2578413523742805e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3440685408326235e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time_median", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2263534835950350e+04, + "cpu_time": 3.2577696285264457e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3439730776384563e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.1234012053865379e-01, + "cpu_time": 8.5487868785279311e+00, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5399703320790699e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.6241474416927541e-05, + "cpu_time": 2.6240648189629199e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6241110214644274e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 44426, + "real_time": 1.5755399022197198e+04, + "cpu_time": 3.7601110183226338e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3191799722328529e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 44426, + "real_time": 1.5757520821214399e+04, + "cpu_time": 3.7634203799577554e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3180597688652496e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 44426, + "real_time": 1.5755654435766673e+04, + "cpu_time": 3.7637409084770457e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3190451107162800e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 44426, + "real_time": 1.5756000223375786e+04, + "cpu_time": 3.7607599648854841e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3188625375582342e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 44426, + "real_time": 1.5757668009459176e+04, + "cpu_time": 3.7604131882230184e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3179820720501757e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5756448502402645e+04, + "cpu_time": 3.7616890919731872e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3186258922845602e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time_median", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5756000223375788e+04, + "cpu_time": 3.7607599648854841e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3188625375582342e+09, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0688743610645428e+00, + "cpu_time": 1.7456254631007344e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6430261158353684e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.7837264273199239e-05, + "cpu_time": 4.6405362602284318e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7836036731369498e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 35952, + "real_time": 1.9468445330877657e+04, + "cpu_time": 4.3319459390297598e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3465071069862478e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 35952, + "real_time": 1.9469018652335417e+04, + "cpu_time": 4.3303269692922571e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3464674551973598e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 35952, + "real_time": 1.9467025964774097e+04, + "cpu_time": 4.3312613540274753e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3466052825652662e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 35952, + "real_time": 1.9470745450162925e+04, + "cpu_time": 4.3304513684914025e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3463480413268229e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 35952, + "real_time": 1.9466708768529650e+04, + "cpu_time": 4.3319535185802764e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3466272245454676e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9468388833335950e+04, + "cpu_time": 4.3311878298842341e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3465110221242329e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time_median", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9468445330877661e+04, + "cpu_time": 4.3312613540274753e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3465071069862478e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6304550827885209e+00, + "cpu_time": 7.8260315360357833e+00, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1276544287679626e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.3748845204728673e-05, + "cpu_time": 1.8069019039160353e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.3746394217330228e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 26402, + "real_time": 2.6509280831779801e+04, + "cpu_time": 5.0500101204454193e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9777526343584324e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 26402, + "real_time": 2.6512112584354589e+04, + "cpu_time": 5.0495233732292458e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9775413910598526e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 26402, + "real_time": 2.6511067602018335e+04, + "cpu_time": 5.0499602189228361e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9776193394795048e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 26402, + "real_time": 2.6512376588185783e+04, + "cpu_time": 5.0545922278613994e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9775216991811619e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 26402, + "real_time": 2.6510264319638707e+04, + "cpu_time": 5.0549358419817276e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9776792629397114e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6511020385195443e+04, + "cpu_time": 5.0518043564881264e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9776228654037323e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time_median", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6511067602018338e+04, + "cpu_time": 5.0500101204454193e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9776193394795048e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2876428271862663e+00, + "cpu_time": 2.7111546952251626e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6054380331143667e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.8570096830574088e-05, + "cpu_time": 5.3667056439808008e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8570625881964675e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 17276, + "real_time": 4.0515799110269116e+04, + "cpu_time": 6.5419956066216742e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5880669344473785e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 17276, + "real_time": 4.0515763606914690e+04, + "cpu_time": 6.5392902002777962e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5880692023315166e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 17276, + "real_time": 4.0511681549578105e+04, + "cpu_time": 6.5430101296598034e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5883299825921940e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 17276, + "real_time": 4.0515556582419849e+04, + "cpu_time": 6.5433753067840829e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5880824267264015e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 17276, + "real_time": 4.0512246085300532e+04, + "cpu_time": 6.5439841166935272e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5882939143689331e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0514209386896466e+04, + "cpu_time": 6.5423310720073772e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5881684920932850e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time_median", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0515556582419849e+04, + "cpu_time": 6.5430101296598034e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5880824267264015e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0616904000191765e+00, + "cpu_time": 1.8468581580990282e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3170951212421979e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.0888081767331497e-05, + "cpu_time": 2.8229359501556967e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0889079488675191e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10289, + "real_time": 6.8334814465553543e+04, + "cpu_time": 9.3945417047330135e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0689364073083714e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10289, + "real_time": 6.8142144219185968e+04, + "cpu_time": 9.3675261249877629e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0776137499493740e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10289, + "real_time": 6.7981925743700311e+04, + "cpu_time": 9.3614845660414183e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0848670099557117e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10289, + "real_time": 6.8327692831926732e+04, + "cpu_time": 9.3993252016719198e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0692562752829948e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10289, + "real_time": 6.8131711486696819e+04, + "cpu_time": 9.3678368548934741e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0780850124533909e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8183657749412669e+04, + "cpu_time": 9.3781428904655186e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0757516909899685e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time_median", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8142144219185968e+04, + "cpu_time": 9.3678368548934741e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0776137499493740e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4892283854347244e+02, + "cpu_time": 1.7421547494814325e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7201043394261673e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1841427030915663e-03, + "cpu_time": 1.8576756292044024e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1848656896172329e-03, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5792, + "real_time": 1.2072106075157732e+05, + "cpu_time": 1.4748171754143783e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4743763630698532e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5792, + "real_time": 1.2084446967688082e+05, + "cpu_time": 1.4771521667817334e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4708282565308220e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5792, + "real_time": 1.2072299147943973e+05, + "cpu_time": 1.4754276968232269e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4743207972230621e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5792, + "real_time": 1.2084516200836700e+05, + "cpu_time": 1.4776004178176296e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4708083718813644e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5792, + "real_time": 1.2072385216040516e+05, + "cpu_time": 1.4761979247237407e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4742960276209961e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2077150721533402e+05, + "cpu_time": 1.4762390763121418e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4729259632652199e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time_median", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2072385216040513e+05, + "cpu_time": 1.4761979247237407e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4742960276209961e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6929395512893478e+01, + "cpu_time": 1.1585354108160281e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9242443004051227e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.5418200083865181e-04, + "cpu_time": 7.8478847322631287e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5407006102599495e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3015, + "real_time": 2.3212904277258393e+05, + "cpu_time": 2.6022232570478087e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6137692637703644e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3015, + "real_time": 2.3215918917174931e+05, + "cpu_time": 2.6028630613599313e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6133000076056358e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3015, + "real_time": 2.3213021463373210e+05, + "cpu_time": 2.6028078009950297e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6137510204072357e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3015, + "real_time": 2.3215351011187062e+05, + "cpu_time": 2.6026375986731265e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6133883980292526e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3015, + "real_time": 2.3212771798842348e+05, + "cpu_time": 2.6022625671640839e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6137898880384254e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3213993493567189e+05, + "cpu_time": 2.6025588570479961e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6135997155701828e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time_median", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3213021463373210e+05, + "cpu_time": 2.6026375986731268e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6137510204072357e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5144208933000286e+01, + "cpu_time": 3.0046990064203175e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3573569948737081e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.5237413533336628e-05, + "cpu_time": 1.1545172161172395e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5235697930691953e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1538, + "real_time": 4.5517874964205874e+05, + "cpu_time": 4.8925332964892109e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6858522093118767e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1538, + "real_time": 4.5520611214378831e+05, + "cpu_time": 4.8927714499350835e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6856306522308945e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1538, + "real_time": 4.5517771081697382e+05, + "cpu_time": 4.8920610858251480e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6858606213137024e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1538, + "real_time": 4.5519882965768158e+05, + "cpu_time": 4.8928543823144294e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6856896166927307e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1538, + "real_time": 4.5516146904403181e+05, + "cpu_time": 4.8933293237969477e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6859921458723015e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5518457426090690e+05, + "cpu_time": 4.8927099076721631e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6858050490843010e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time_median", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5517874964205874e+05, + "cpu_time": 4.8927714499350835e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6858522093118767e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7902377917613968e+01, + "cpu_time": 4.6381516000840733e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4496247238509697e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.9329931043209125e-05, + "cpu_time": 9.4797191895866907e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9329934832313321e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 776, + "real_time": 9.0165870132455218e+05, + "cpu_time": 9.4742139948441042e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7214116550650444e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 776, + "real_time": 9.0166447483442398e+05, + "cpu_time": 9.4743642783495155e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7213878262378838e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 776, + "real_time": 9.0166405837961298e+05, + "cpu_time": 9.4741972036072961e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7213895450486198e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 776, + "real_time": 9.0165725588431279e+05, + "cpu_time": 9.4742789304114971e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7214176208332100e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 776, + "real_time": 9.0167354480315419e+05, + "cpu_time": 9.4745992525775405e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7213503926551735e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0166360704521113e+05, + "cpu_time": 9.4743307319579925e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7213914079679863e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time_median", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0166405837961286e+05, + "cpu_time": 9.4742789304114971e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7213895450486198e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.4054533835177025e+00, + "cpu_time": 1.6383635801129611e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6436756956934032e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.1040389491915258e-06, + "cpu_time": 1.7292657671180671e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1039979563368352e-06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 390, + "real_time": 1.7947245159377465e+06, + "cpu_time": 1.8645698743590889e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7392292468315407e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 390, + "real_time": 1.7946711177221285e+06, + "cpu_time": 1.8644369051283041e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7393405029650986e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 390, + "real_time": 1.7946739925835759e+06, + "cpu_time": 1.8645012333335076e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7393345129714310e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 390, + "real_time": 1.7946776856357853e+06, + "cpu_time": 1.8644881307692137e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7393268182429039e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 390, + "real_time": 1.7947157859826125e+06, + "cpu_time": 1.8646701384615228e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7392474353958885e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7946926195723698e+06, + "cpu_time": 1.8645332564103275e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7392957032813728e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time_median", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7946776856357851e+06, + "cpu_time": 1.8645012333335076e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7393268182429039e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5428213735452722e+01, + "cpu_time": 9.0031859627769705e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2980126205965190e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4168562046859939e-05, + "cpu_time": 4.8286539978966412e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4168477277545377e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 195, + "real_time": 3.5809334725714647e+06, + "cpu_time": 3.7104005487184948e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481212378854500e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 195, + "real_time": 3.5809064021286294e+06, + "cpu_time": 3.7102641897438215e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481495724159615e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 195, + "real_time": 3.5808358257875228e+06, + "cpu_time": 3.7098207948713773e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7482234464206940e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 195, + "real_time": 3.5809506889050589e+06, + "cpu_time": 3.7103595999998362e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481032178368118e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 195, + "real_time": 3.5808858819879019e+06, + "cpu_time": 3.7101362615384208e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481710510553894e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5809024542761156e+06, + "cpu_time": 3.7101962789743901e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481537051228615e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time_median", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5809064021286294e+06, + "cpu_time": 3.7102641897438215e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481495724159615e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4772437669193451e+01, + "cpu_time": 2.3323923628735528e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6863874905944342e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2503115692450288e-05, + "cpu_time": 6.2864392810999656e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2503189194693973e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 98, + "real_time": 7.1530135357942507e+06, + "cpu_time": 7.3861540612242334e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7527603527761765e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 98, + "real_time": 7.1529531233696910e+06, + "cpu_time": 7.3860815816320907e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7527920478464211e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 98, + "real_time": 7.1530951690689037e+06, + "cpu_time": 7.3863855102039063e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7527175251457115e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 98, + "real_time": 7.1528486346787941e+06, + "cpu_time": 7.3859355918371798e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7528468685686699e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 98, + "real_time": 7.1528763985451385e+06, + "cpu_time": 7.3858466428574435e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7528323019058250e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1529573722913563e+06, + "cpu_time": 7.3860806775509715e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7527898192485619e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time_median", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1529531233696919e+06, + "cpu_time": 7.3860815816320907e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7527920478464211e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0067390775274653e+02, + "cpu_time": 2.0864502088968885e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2818203892218822e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4074445367552494e-05, + "cpu_time": 2.8248408052708953e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4074383708170166e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49, + "real_time": 1.4297099650970528e+07, + "cpu_time": 1.4720706857142616e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551036581294006e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49, + "real_time": 1.4296812347459549e+07, + "cpu_time": 1.4720467816325290e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551791193188492e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49, + "real_time": 1.4297027768073034e+07, + "cpu_time": 1.4720667795918232e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551225381187042e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49, + "real_time": 1.4296629542142762e+07, + "cpu_time": 1.4719948102041673e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7552271353009712e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49, + "real_time": 1.4296616408593800e+07, + "cpu_time": 1.4720305285714969e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7552305850304756e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4296837143447936e+07, + "cpu_time": 1.4720419171428556e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551726071796806e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time_median", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4296812347459551e+07, + "cpu_time": 1.4720467816325292e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551791193188492e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2230878368830494e+02, + "cpu_time": 3.0895786428595409e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8391013007140066e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5549508010601237e-05, + "cpu_time": 2.0988387673472137e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5549488429772762e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 24, + "real_time": 2.8585107919449609e+07, + "cpu_time": 2.9388081958330762e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7562979542554565e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 24, + "real_time": 2.8585060965269804e+07, + "cpu_time": 2.9387933333334595e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7563041243976067e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 24, + "real_time": 2.8584861118967336e+07, + "cpu_time": 2.9388330291666877e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7563303859731689e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 24, + "real_time": 2.8585374355316162e+07, + "cpu_time": 2.9388557791667342e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7562629429070641e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 24, + "real_time": 2.8585894654194515e+07, + "cpu_time": 2.9389574333336551e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7561945742441399e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8585259802639484e+07, + "cpu_time": 2.9388495541667223e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7562779963554878e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time_median", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8585107919449609e+07, + "cpu_time": 2.9388330291666877e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7562979542554565e+10, + "own_gpu": 3.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9928460399800508e+02, + "cpu_time": 6.4847967017841904e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2467854204264923e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3968199231169354e-05, + "cpu_time": 2.2065766151894361e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3968043434264351e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57080, + "real_time": 1.2261260430753995e+04, + "cpu_time": 3.3466426769445090e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3406027244363165e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57080, + "real_time": 1.2264207094591759e+04, + "cpu_time": 3.3403966590749093e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3398000934004486e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57080, + "real_time": 1.2263321058495751e+04, + "cpu_time": 3.3506142466712598e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3400413969936669e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57080, + "real_time": 1.2261616701992434e+04, + "cpu_time": 3.3458255693763873e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3405056605092102e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57080, + "real_time": 1.2264576575842399e+04, + "cpu_time": 3.3634982568323830e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3396994789595205e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2262996372335268e+04, + "cpu_time": 3.3493954817798898e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3401298708598328e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time_median", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2263321058495752e+04, + "cpu_time": 3.3466426769445083e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3400413969936669e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4987741705402411e+00, + "cpu_time": 8.6843921943942718e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0823594403236959e+04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2221924601734399e-04, + "cpu_time": 2.5928237622686858e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2222157814698368e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57081, + "real_time": 1.2262265415366213e+04, + "cpu_time": 3.3627538603038796e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6806578739800870e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57081, + "real_time": 1.2260440714283041e+04, + "cpu_time": 3.3563467370928978e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6816521452255535e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57081, + "real_time": 1.2263144969048621e+04, + "cpu_time": 3.3563416478336221e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6801787149023151e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57081, + "real_time": 1.2260672591328328e+04, + "cpu_time": 3.3634224452970288e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6815257800734353e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57081, + "real_time": 1.2259723206693387e+04, + "cpu_time": 3.3590401271877461e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6820431928899097e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2261249379343919e+04, + "cpu_time": 3.3595809635430356e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6812115414142609e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time_median", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2260672591328326e+04, + "cpu_time": 3.3590401271877468e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6815257800734353e+08, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4090259136127774e+00, + "cpu_time": 3.3937496544479316e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6775647701598718e+04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1491699336827063e-04, + "cpu_time": 1.0101705216441223e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1491276279114350e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57071, + "real_time": 1.2263166340044489e+04, + "cpu_time": 3.3963914159556312e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3360334146736004e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57071, + "real_time": 1.2262319483218602e+04, + "cpu_time": 3.3922213383328257e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3361256834338770e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57071, + "real_time": 1.2261537826315061e+04, + "cpu_time": 3.3978376811340248e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3362108596882138e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57071, + "real_time": 1.2261290521540757e+04, + "cpu_time": 3.4039529673563251e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3362378104667227e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57071, + "real_time": 1.2261635351952569e+04, + "cpu_time": 3.4011537768743729e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3362002318386495e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2261989904614295e+04, + "cpu_time": 3.3983114359306361e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3361616000202129e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time_median", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2261635351952567e+04, + "cpu_time": 3.3978376811340248e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3362002318386495e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6023703650054464e-01, + "cpu_time": 4.4973877954651769e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2837861874869748e+04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.1999483151952407e-05, + "cpu_time": 1.3234183741707461e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1996888605103312e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57086, + "real_time": 1.2264711294775700e+04, + "cpu_time": 3.4484693742772804e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6717302358317981e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57086, + "real_time": 1.2263937395814692e+04, + "cpu_time": 3.4382277108222515e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6718988317065873e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57086, + "real_time": 1.2264811948735653e+04, + "cpu_time": 3.4418894685211497e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6717083096718798e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57086, + "real_time": 1.2263997236809302e+04, + "cpu_time": 3.4454998563572080e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6718857944332991e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57086, + "real_time": 1.2265588813265000e+04, + "cpu_time": 3.4458759065270744e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6715390919155903e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2264609337880069e+04, + "cpu_time": 3.4439924633009927e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6717524527118311e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time_median", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2264711294775701e+04, + "cpu_time": 3.4454998563572080e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6717302358317981e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7767842518035437e-01, + "cpu_time": 3.9848865192542419e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4762440123502619e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.5254790960793102e-05, + "cpu_time": 1.1570543669061381e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5253772139401348e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57122, + "real_time": 1.2253964631494018e+04, + "cpu_time": 3.5281728826021135e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3481466587201805e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57122, + "real_time": 1.2252925193353705e+04, + "cpu_time": 3.5219073754419805e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3486003518203449e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57122, + "real_time": 1.2250869435134899e+04, + "cpu_time": 3.5227314589825481e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3494978741709490e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57122, + "real_time": 1.2251511444566187e+04, + "cpu_time": 3.5219100591716211e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3492175472820244e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57122, + "real_time": 1.2253808193994926e+04, + "cpu_time": 3.5239633836349014e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3482149355101233e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2252615779708749e+04, + "cpu_time": 3.5237370319666334e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3487354735007248e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time_median", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2252925193353705e+04, + "cpu_time": 3.5227314589825481e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3486003518203449e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3789729168077893e+00, + "cpu_time": 2.6184277504215817e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0199053281592391e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1254518558326721e-04, + "cpu_time": 7.4308262128181880e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1254819682116820e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 44488, + "real_time": 1.5733384244771998e+04, + "cpu_time": 4.0322709629565405e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3308204999540091e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 44488, + "real_time": 1.5731974807325549e+04, + "cpu_time": 4.0266409121561825e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3315668633645859e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 44488, + "real_time": 1.5731129476158418e+04, + "cpu_time": 4.0400497819637938e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3320145701329594e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 44488, + "real_time": 1.5733265740094212e+04, + "cpu_time": 4.0354951582451075e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3308832486048851e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 44488, + "real_time": 1.5733014659922552e+04, + "cpu_time": 4.0409858770905441e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3310161995771780e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5732553785654547e+04, + "cpu_time": 4.0350885384824338e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3312602763267231e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time_median", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5733014659922552e+04, + "cpu_time": 4.0354951582451075e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.3310161995771780e+09, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7118740433920669e-01, + "cpu_time": 5.8919224697651430e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1431543441743997e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.1731071609287411e-05, + "cpu_time": 1.4601717938960144e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1733209305543763e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 35977, + "real_time": 1.9455984883557852e+04, + "cpu_time": 4.5984581093476023e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3473694678984690e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 35977, + "real_time": 1.9454792625772730e+04, + "cpu_time": 4.5870060205132191e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3474520394153410e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 35977, + "real_time": 1.9454094593126858e+04, + "cpu_time": 4.5962629819048969e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3475003873612068e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 35977, + "real_time": 1.9456513242316149e+04, + "cpu_time": 4.6037705422908504e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3473328788935349e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 35977, + "real_time": 1.9454925463418487e+04, + "cpu_time": 4.5965871306668720e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3474428390533541e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9455262161638417e+04, + "cpu_time": 4.5964169569446887e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3474195225243813e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time_median", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9454925463418487e+04, + "cpu_time": 4.5965871306668720e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3474428390533541e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.7269924831009558e-01, + "cpu_time": 6.0592885825319378e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7366002962918917e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.9996717609287674e-05, + "cpu_time": 1.3182634733293742e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9996309120346698e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 26436, + "real_time": 2.6477830882206774e+04, + "cpu_time": 5.3052355651386948e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9801017777189747e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 26436, + "real_time": 2.6475434148644832e+04, + "cpu_time": 5.3037153124525619e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9802810297894062e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 26436, + "real_time": 2.6477728112677029e+04, + "cpu_time": 5.3004312036618081e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9801094632019466e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 26436, + "real_time": 2.6477073736881299e+04, + "cpu_time": 5.2977924534724276e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9801584012272923e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 26436, + "real_time": 2.6475554886509151e+04, + "cpu_time": 5.2988329172337886e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9802719990097565e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6476724353383819e+04, + "cpu_time": 5.3012014903918564e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9801845341894753e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time_median", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6477073736881302e+04, + "cpu_time": 5.3004312036618081e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9801584012272923e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1604189728995216e+00, + "cpu_time": 3.1787605125958020e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.6788129695252678e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.3827890392008258e-05, + "cpu_time": 5.9963020050400552e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3828303977122317e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 17285, + "real_time": 4.0499767597695965e+04, + "cpu_time": 6.7159375990744375e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5890914002668339e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 17285, + "real_time": 4.0495428006946466e+04, + "cpu_time": 6.7049465490309609e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5893688537385761e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 17285, + "real_time": 4.0493825183149725e+04, + "cpu_time": 6.7088463754701530e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5894713459580326e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 17285, + "real_time": 4.0496593906015718e+04, + "cpu_time": 6.7113299971074928e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5892943056730392e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 17285, + "real_time": 4.0492557072135089e+04, + "cpu_time": 6.7105251374020518e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5895524407905975e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0495634353188594e+04, + "cpu_time": 6.7103171316170206e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5893556692854160e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time_median", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0495428006946466e+04, + "cpu_time": 6.7105251374020518e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.5893688537385761e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7744272732519630e+00, + "cpu_time": 3.9895205939404136e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7739613697710556e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.8511762256997629e-05, + "cpu_time": 5.9453532756938989e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8509760586911396e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10286, + "real_time": 6.8078314875118624e+04, + "cpu_time": 9.4875103344345247e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0804992806402004e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10286, + "real_time": 6.8191617501816931e+04, + "cpu_time": 9.5045485125409192e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0753809292529575e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10286, + "real_time": 6.8213138921454243e+04, + "cpu_time": 9.5196397141746420e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0744106387111420e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10286, + "real_time": 6.8234924239211236e+04, + "cpu_time": 9.5123244215434985e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0734290737218559e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10286, + "real_time": 6.8080320686723760e+04, + "cpu_time": 9.4931478806143146e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0804085216492855e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8159663244864976e+04, + "cpu_time": 9.5034341726615807e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0768256887950886e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time_median", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8191617501816931e+04, + "cpu_time": 9.5045485125409177e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 3.0753809292529575e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4929531204256250e+01, + "cpu_time": 1.3250229075015477e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3833695960518412e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0993236708795112e-03, + "cpu_time": 1.3942569427304772e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0996299232592529e-03, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5794, + "real_time": 1.2073241482629217e+05, + "cpu_time": 1.4843050949258605e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4740496212510086e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5794, + "real_time": 1.2081014825032407e+05, + "cpu_time": 1.4856975578183771e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4718142976773880e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5794, + "real_time": 1.2072965329522437e+05, + "cpu_time": 1.4848015913012842e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4741290855391792e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5794, + "real_time": 1.2081296236820599e+05, + "cpu_time": 1.4863603658958140e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4717334280876831e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5794, + "real_time": 1.2073152597275631e+05, + "cpu_time": 1.4857569209525562e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4740751980112190e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2076334094256058e+05, + "cpu_time": 1.4853843061787786e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4731603261132957e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time_median", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2073241482629217e+05, + "cpu_time": 1.4856975578183771e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.4740496212510086e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4036020746693382e+01, + "cpu_time": 8.2065743221675888e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2663097744131962e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6464725472971563e-04, + "cpu_time": 5.5248828791515836e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6459870996807207e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3016, + "real_time": 2.3213452258501176e+05, + "cpu_time": 2.6187314588859299e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6136839564343315e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3016, + "real_time": 2.3208592761243417e+05, + "cpu_time": 2.6185974535811110e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6144406023653175e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3016, + "real_time": 2.3214040961635634e+05, + "cpu_time": 2.6197133720158800e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6135923141788704e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3016, + "real_time": 2.3208629211475109e+05, + "cpu_time": 2.6200249966840190e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6144349257182312e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3016, + "real_time": 2.3213807281326110e+05, + "cpu_time": 2.6197573872679184e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6136286901752869e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3211704494836292e+05, + "cpu_time": 2.6193649336869718e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6139560977744072e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time_median", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3213452258501179e+05, + "cpu_time": 2.6197133720158800e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.6136839564343315e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8317715095299160e+01, + "cpu_time": 6.5219404324282337e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4091270991705377e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2199756851806707e-04, + "cpu_time": 2.4898937710248967e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2200278531014317e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1538, + "real_time": 4.5517738917990861e+05, + "cpu_time": 4.9009228933679679e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6858632258134453e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1538, + "real_time": 4.5516854742486199e+05, + "cpu_time": 4.9008389271775843e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6859348245650780e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1538, + "real_time": 4.5518248715672299e+05, + "cpu_time": 4.9018049739923287e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6858219446881912e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1538, + "real_time": 4.5516928221142682e+05, + "cpu_time": 4.9010239271785005e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6859288743054855e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1538, + "real_time": 4.5518748591937392e+05, + "cpu_time": 4.9018367620288976e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6857814678525017e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5517703837845894e+05, + "cpu_time": 4.9012854967490555e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6858660674449402e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time_median", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5517738917990867e+05, + "cpu_time": 4.9010239271785005e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.6858632258134453e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.2343433709861049e+00, + "cpu_time": 4.9322262714033982e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6678761740152317e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8090419060505475e-05, + "cpu_time": 1.0063127876706764e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8090391924190114e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 776, + "real_time": 9.0167600587706466e+05, + "cpu_time": 9.4810044329899934e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7213402354387192e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 776, + "real_time": 9.0157517595648684e+05, + "cpu_time": 9.4791530283514143e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7217564208555199e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 776, + "real_time": 9.0167518616918009e+05, + "cpu_time": 9.4819675773191871e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7213436184883797e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 776, + "real_time": 9.0153187763284333e+05, + "cpu_time": 9.4801565721656953e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7219351675177628e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 776, + "real_time": 9.0167930736158462e+05, + "cpu_time": 9.4810102448442671e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7213266098103165e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0162751059943193e+05, + "cpu_time": 9.4806583711341117e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7215404104221390e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time_median", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0167518616918009e+05, + "cpu_time": 9.4810044329899922e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7213436184883797e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9268021200381739e+01, + "cpu_time": 1.0577634214086231e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8592022168989726e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.6825540909161087e-05, + "cpu_time": 1.1157067157162943e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6828460840887386e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 390, + "real_time": 1.7946201666162754e+06, + "cpu_time": 1.8652801282049343e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7394466666744621e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 390, + "real_time": 1.7946710424999206e+06, + "cpu_time": 1.8651841025641074e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7393406596965790e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 390, + "real_time": 1.7946000581081861e+06, + "cpu_time": 1.8653105179489616e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7394885672044479e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 390, + "real_time": 1.7946025204414930e+06, + "cpu_time": 1.8652506615382708e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7394834363372253e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 390, + "real_time": 1.7946295956603417e+06, + "cpu_time": 1.8652857538462251e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7394270194962990e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7946246766652435e+06, + "cpu_time": 1.8652622328204997e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7394372698818024e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time_median", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7946201666162754e+06, + "cpu_time": 1.8652801282049343e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.7394466666744621e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8684919785585965e+01, + "cpu_time": 4.8587976406616477e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9769695599024091e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5983798817972371e-05, + "cpu_time": 2.6048871601901055e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5983606966861438e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 195, + "real_time": 3.5809293198279846e+06, + "cpu_time": 3.7112841538461135e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481255845185837e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 195, + "real_time": 3.5809227301237676e+06, + "cpu_time": 3.7111114717950281e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481324819137054e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 195, + "real_time": 3.5809711469576145e+06, + "cpu_time": 3.7115209076925265e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7480818049603973e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 195, + "real_time": 3.5810384457596601e+06, + "cpu_time": 3.7112968461536719e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7480113668963379e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 195, + "real_time": 3.5808997622762737e+06, + "cpu_time": 3.7115032820513360e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481565223898285e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5809522809890606e+06, + "cpu_time": 3.7113433323077350e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481015521357712e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time_median", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5809293198279841e+06, + "cpu_time": 3.7112968461536719e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.7481255845185837e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4632798121335497e+01, + "cpu_time": 1.7068925425636652e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7182257611955132e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5256499901262549e-05, + "cpu_time": 4.5991232546580625e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5256325586848390e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 98, + "real_time": 7.1529610871280339e+06, + "cpu_time": 7.3871267959185494e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7527878696706123e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 98, + "real_time": 7.1530623304448565e+06, + "cpu_time": 7.3867885204080902e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7527347533025864e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 98, + "real_time": 7.1530019180202968e+06, + "cpu_time": 7.3868247040817253e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7527664479404144e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 98, + "real_time": 7.1528761799694328e+06, + "cpu_time": 7.3870431734687779e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7528324165838860e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 98, + "real_time": 7.1527537538156826e+06, + "cpu_time": 7.3865333469389407e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7528966498644157e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1529310538756605e+06, + "cpu_time": 7.3868633081632182e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7528036274723839e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time_median", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1529610871280329e+06, + "cpu_time": 7.3868247040817244e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.7527878696706123e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2000386942569791e+02, + "cpu_time": 2.3331052692737177e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2960752625742962e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6776880487429894e-05, + "cpu_time": 3.1584519327647564e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6776990984777094e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 49, + "real_time": 1.4296756221019492e+07, + "cpu_time": 1.4720904367346074e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551938614626259e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 49, + "real_time": 1.4296965160388120e+07, + "cpu_time": 1.4721432306122996e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551389821350418e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 49, + "real_time": 1.4297036967259280e+07, + "cpu_time": 1.4721083285714522e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551201219487183e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 49, + "real_time": 1.4296870488597423e+07, + "cpu_time": 1.4721003734693186e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551638481175690e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 49, + "real_time": 1.4296811682229139e+07, + "cpu_time": 1.4721384448978137e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551792940472710e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4296888103898693e+07, + "cpu_time": 1.4721161628570985e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551592215422455e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time_median", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4296870488597419e+07, + "cpu_time": 1.4721083285714522e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7551638481175690e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1362376407920799e+02, + "cpu_time": 2.3460955131261366e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9843880069454777e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.9474472524005646e-06, + "cpu_time": 1.5936891206824397e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.9474339991362288e-06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 24, + "real_time": 2.8586001368239522e+07, + "cpu_time": 2.9390131999998439e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7561805520410454e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 24, + "real_time": 2.8586894972249866e+07, + "cpu_time": 2.9391024958333865e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7560631367705818e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 24, + "real_time": 2.8584967677791912e+07, + "cpu_time": 2.9388645083334569e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7563163831533943e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 24, + "real_time": 2.8585627830276888e+07, + "cpu_time": 2.9389113041664436e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7562296353090088e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 24, + "real_time": 2.8585280912617844e+07, + "cpu_time": 2.9389290458330680e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7562752217909431e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8585754552235208e+07, + "cpu_time": 2.9389641108332403e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7562129858129944e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time_median", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8585627830276888e+07, + "cpu_time": 2.9389290458330683e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7562296353090088e+10, + "own_gpu": 4.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4515773364450092e+02, + "cpu_time": 9.4205757977892677e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.7913524491767737e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6067450214856572e-05, + "cpu_time": 3.2054068857337606e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6067085349414856e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57030, + "real_time": 1.2250648481064340e+04, + "cpu_time": 3.1200915360337596e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3434964739467716e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57030, + "real_time": 1.2261579275654864e+04, + "cpu_time": 3.1172345204277972e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3405158568215847e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57030, + "real_time": 1.2264264368075752e+04, + "cpu_time": 3.1213434543221858e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3397844967057389e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57030, + "real_time": 1.2252449458469930e+04, + "cpu_time": 3.1194847518851140e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3430050161671942e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57030, + "real_time": 1.2257444071804395e+04, + "cpu_time": 3.1201317482025388e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3416428221132690e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2257277131013858e+04, + "cpu_time": 3.1196572021742795e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3416889331509119e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time_median", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2257444071804395e+04, + "cpu_time": 3.1200915360337600e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3416428221132690e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.8009314339842728e+00, + "cpu_time": 1.5129781954491600e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5814859076197929e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.7326427982170047e-04, + "cpu_time": 4.8498219432400232e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7325946228292230e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57241, + "real_time": 1.2252486989997618e+04, + "cpu_time": 3.1204193008506416e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6859895519069576e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57241, + "real_time": 1.2255875393931577e+04, + "cpu_time": 3.1136875229294037e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6841410643390024e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57241, + "real_time": 1.2254429806069651e+04, + "cpu_time": 3.1198086266836173e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6849295557941687e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57241, + "real_time": 1.2251700876519912e+04, + "cpu_time": 3.1190809996330456e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6864185491989684e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57241, + "real_time": 1.2246667323272946e+04, + "cpu_time": 3.1205516919690137e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6891667616644883e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2252232077958342e+04, + "cpu_time": 3.1187096284131450e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6861290965807176e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time_median", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2252486989997618e+04, + "cpu_time": 3.1198086266836173e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6859895519069576e+08, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5148758868507555e+00, + "cpu_time": 2.8672040060627353e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9184521740194620e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8687637195299188e-04, + "cpu_time": 9.1935587075530961e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8693017234748835e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57244, + "real_time": 1.2228349285219010e+04, + "cpu_time": 3.1467156732583404e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3398374235027883e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57244, + "real_time": 1.2227971308188162e+04, + "cpu_time": 3.1280754803997843e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3398788390212247e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57244, + "real_time": 1.2227322963409306e+04, + "cpu_time": 3.1466993781009103e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3399498851081054e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57244, + "real_time": 1.2224486685783482e+04, + "cpu_time": 3.1345997344700536e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3402607750437357e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57244, + "real_time": 1.2226823309924344e+04, + "cpu_time": 3.1454189574452579e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3400046426369252e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2226990710504862e+04, + "cpu_time": 3.1403018447348695e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3399863130625558e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time_median", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2227322963409308e+04, + "cpu_time": 3.1454189574452583e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3399498851081054e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5178810640734641e+00, + "cpu_time": 8.5183407760805665e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6636640333913575e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2414183506080293e-04, + "cpu_time": 2.7125866229587734e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2415530048131852e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57186, + "real_time": 1.2243312071067616e+04, + "cpu_time": 3.1912615028153199e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6763999651233778e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57186, + "real_time": 1.2237201269873654e+04, + "cpu_time": 3.1757756706187021e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6777364592891364e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57186, + "real_time": 1.2243586316959874e+04, + "cpu_time": 3.1870052950021924e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6763400160465736e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57186, + "real_time": 1.2241778816141485e+04, + "cpu_time": 3.1821885286610108e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6767351781257086e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57186, + "real_time": 1.2244803484545742e+04, + "cpu_time": 3.1935120256705453e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6760739803914971e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2242136391717677e+04, + "cpu_time": 3.1859486045535545e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6766571197952590e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time_median", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2243312071067619e+04, + "cpu_time": 3.1870052950021927e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6763999651233778e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9614682632210427e+00, + "cpu_time": 7.1445009832680668e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4764992716744752e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4190779848071303e-04, + "cpu_time": 2.2425035272247351e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4196223056653109e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57027, + "real_time": 1.2273278884209927e+04, + "cpu_time": 3.2718596629665095e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3397303702040644e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57027, + "real_time": 1.2276864389541019e+04, + "cpu_time": 3.2609850228838324e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3381708814696875e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57027, + "real_time": 1.2272978130593869e+04, + "cpu_time": 3.2684937485753333e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3398612221619616e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57027, + "real_time": 1.2276555678472298e+04, + "cpu_time": 3.2672538692199083e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3383051172016792e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57027, + "real_time": 1.2272771280635161e+04, + "cpu_time": 3.2737367334772855e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3399512222155800e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2274489672690455e+04, + "cpu_time": 3.2684658074245737e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3392037626505947e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time_median", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2273278884209927e+04, + "cpu_time": 3.2684937485753337e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3397303702040644e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0378480904035934e+00, + "cpu_time": 4.9176373686243238e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8637907845345722e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6602304004032092e-04, + "cpu_time": 1.5045705411552812e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6601334540815933e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 43925, + "real_time": 1.5935099115920548e+04, + "cpu_time": 3.8032167125781612e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.2253645896088390e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 43925, + "real_time": 1.5932053307819679e+04, + "cpu_time": 3.7915072373363277e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.2269370725534792e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 43925, + "real_time": 1.5936510336413105e+04, + "cpu_time": 3.7991421627773969e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.2246362116375914e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 43925, + "real_time": 1.5935299164917311e+04, + "cpu_time": 3.7998262219692886e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.2252613298007164e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 43925, + "real_time": 1.5935536383021381e+04, + "cpu_time": 3.8024403665338978e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.2251388876782017e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5934899661618405e+04, + "cpu_time": 3.7992265402390149e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.2254676182557669e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time_median", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5935299164917309e+04, + "cpu_time": 3.7998262219692886e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 8.2252613298007164e+09, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6808332975488018e+00, + "cpu_time": 4.6423197267075025e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.6772207458379201e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0548126020506679e-04, + "cpu_time": 1.2219117963982867e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0549212699566798e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 35258, + "real_time": 1.9850857906383819e+04, + "cpu_time": 4.4163758551250925e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3205676109126617e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 35258, + "real_time": 1.9849641616455287e+04, + "cpu_time": 4.3862553831755642e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3206485289018190e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 35258, + "real_time": 1.9853580822497464e+04, + "cpu_time": 4.4096904787564265e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3203864952308578e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 35258, + "real_time": 1.9854274513592365e+04, + "cpu_time": 4.3988119717511734e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3203403620743460e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 35258, + "real_time": 1.9852283023240001e+04, + "cpu_time": 4.4124716036076228e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3204728125884670e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9852127576433788e+04, + "cpu_time": 4.4047210584831759e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3204831619416304e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time_median", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9852283023240001e+04, + "cpu_time": 4.4096904787564265e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.3204728125884670e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9064027664361280e+00, + "cpu_time": 1.2211662897961469e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2680814711997018e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.6030148864205102e-05, + "cpu_time": 2.7724032318556663e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6031627494221327e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 25675, + "real_time": 2.7264763213224356e+04, + "cpu_time": 5.1517422122690325e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9229508648206493e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 25675, + "real_time": 2.7262780662914971e+04, + "cpu_time": 5.1341946368059340e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9230907018709900e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 25675, + "real_time": 2.7270093180905496e+04, + "cpu_time": 5.1513697487830243e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9225750221018902e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 25675, + "real_time": 2.7268451392384108e+04, + "cpu_time": 5.1415033768258771e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9226907771756706e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 25675, + "real_time": 2.7266596589966211e+04, + "cpu_time": 5.1488185589092624e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9228215676647076e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7266537007879029e+04, + "cpu_time": 5.1455257067186256e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9228257867267818e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time_median", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7266596589966204e+04, + "cpu_time": 5.1488185589092624e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.9228215676647076e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8970355501924621e+00, + "cpu_time": 7.5533589812838187e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0429931397241645e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0624875279744270e-04, + "cpu_time": 1.4679469915816828e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0624951848612053e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 16637, + "real_time": 4.2075368417441423e+04, + "cpu_time": 6.7560925226906984e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.4921374177803650e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 16637, + "real_time": 4.2075544173780196e+04, + "cpu_time": 6.7425998857966595e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.4921270077201542e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 16637, + "real_time": 4.2079650145093685e+04, + "cpu_time": 6.7506499068341276e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.4918838354987125e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 16637, + "real_time": 4.2078764538335578e+04, + "cpu_time": 6.7446936707340967e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.4919362806973618e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 16637, + "real_time": 4.2077139885268465e+04, + "cpu_time": 6.7551634489397242e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.4920324975964314e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2077293431983875e+04, + "cpu_time": 6.7498398869990633e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.4920234078586048e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time_median", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2077139885268472e+04, + "cpu_time": 6.7506499068341276e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 2.4920324975964314e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9044298339085466e+00, + "cpu_time": 6.0620346463602097e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1278897058134719e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.5260274095028830e-05, + "cpu_time": 8.9810051021156211e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5259996445325017e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9826, + "real_time": 7.1267099252624219e+04, + "cpu_time": 9.7091479442302239e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.9426650193325752e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9826, + "real_time": 7.1179719086906669e+04, + "cpu_time": 9.6884420008145884e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.9462774325359283e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9826, + "real_time": 7.1247687342773308e+04, + "cpu_time": 9.7110829737439097e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.9434667681360401e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9826, + "real_time": 7.1276736262186605e+04, + "cpu_time": 9.7200397313258247e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.9422671547218010e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9826, + "real_time": 7.1184832537505994e+04, + "cpu_time": 9.6831355790758986e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.9460657913257694e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1231214896399368e+04, + "cpu_time": 9.7023696458380902e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.9441484332104233e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time_median", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1247687342773308e+04, + "cpu_time": 9.7091479442302254e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.9434667681360401e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5919399125050106e+01, + "cpu_time": 1.5795790210731144e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8982421487243403e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.4465275780901021e-04, + "cpu_time": 1.6280342624862656e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4475083094041460e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5510, + "real_time": 1.2704977420130796e+05, + "cpu_time": 1.5385468874773654e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3013077168907085e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5510, + "real_time": 1.2704495616517487e+05, + "cpu_time": 1.5372612105263007e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3014329152484119e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5510, + "real_time": 1.2704783926139875e+05, + "cpu_time": 1.5379821796734206e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3013579958414654e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5510, + "real_time": 1.2705129047654173e+05, + "cpu_time": 1.5382861270418635e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3012683179116707e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5510, + "real_time": 1.2703029450579308e+05, + "cpu_time": 1.5384224627950561e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3018139620299183e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2704483092204327e+05, + "cpu_time": 1.5380997735028013e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3014361815844357e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time_median", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2704783926139874e+05, + "cpu_time": 1.5382861270418635e+05, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.3013579958414654e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.4636729506510271e+00, + "cpu_time": 5.1375653319661367e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1995631508097239e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.6619577429674964e-05, + "cpu_time": 3.3402029052160051e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6624433423217121e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2863, + "real_time": 2.4447221017562677e+05, + "cpu_time": 2.7253116032133502e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4313135198367516e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2863, + "real_time": 2.4451144421284474e+05, + "cpu_time": 2.7253745930838917e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4307629350460182e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2863, + "real_time": 2.4450384201379470e+05, + "cpu_time": 2.7256295214809489e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4308696055282116e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2863, + "real_time": 2.4451809669484539e+05, + "cpu_time": 2.7269972651064763e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4306695959885723e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2863, + "real_time": 2.4446685173653829e+05, + "cpu_time": 2.7257093782746146e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4313887303790356e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4449448896673002e+05, + "cpu_time": 2.7258044722318568e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4310008773557178e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time_median", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4450384201379470e+05, + "cpu_time": 2.7256295214809489e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.4308696055282116e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3411779538857211e+01, + "cpu_time": 6.8740843339152292e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2854660156757063e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.5755857883745644e-05, + "cpu_time": 2.5218552555557328e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.5758238867249264e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1458, + "real_time": 4.8018269090266799e+05, + "cpu_time": 5.1399065637854621e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4939235249112099e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1458, + "real_time": 4.8012451386265713e+05, + "cpu_time": 5.1385185665296519e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4943468861911179e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1458, + "real_time": 4.8014417513697164e+05, + "cpu_time": 5.1401618381345132e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4942037972685875e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1458, + "real_time": 4.8006965360136010e+05, + "cpu_time": 5.1389469478739740e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4947462048770638e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1458, + "real_time": 4.8014844094408833e+05, + "cpu_time": 5.1401153909465432e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4941727535367859e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8013389488954906e+05, + "cpu_time": 5.1395298614540289e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4942786333569534e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time_median", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8014417513697158e+05, + "cpu_time": 5.1399065637854626e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4942037972685875e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1573033897529783e+01, + "cpu_time": 7.4944139545005456e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0257005402650144e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.6586334228899467e-05, + "cpu_time": 1.4581905653876859e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.6590133694010082e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 736, + "real_time": 9.5107386218644585e+05, + "cpu_time": 9.9687406657607097e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.5280574237274208e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 736, + "real_time": 9.5102841829509044e+05, + "cpu_time": 9.9673068614120316e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.5282260082357018e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 736, + "real_time": 9.5117604091090325e+05, + "cpu_time": 9.9699780434786808e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.5276784272095695e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 736, + "real_time": 9.5109646423699579e+05, + "cpu_time": 9.9694280978252192e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.5279735822505226e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 736, + "real_time": 9.5116298641938635e+05, + "cpu_time": 9.9699648505428527e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.5277268437782959e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5110755440976447e+05, + "cpu_time": 9.9690837038038985e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.5279324570403023e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time_median", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5109646423699579e+05, + "cpu_time": 9.9694280978252203e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.5279735822505226e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1813813177313321e+01, + "cpu_time": 1.1146427981695331e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2928596804514662e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.4991401751270464e-05, + "cpu_time": 1.1180995478493369e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4991598007378549e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 371, + "real_time": 1.8867470419430989e+06, + "cpu_time": 1.9565940673855017e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5568553975781929e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 371, + "real_time": 1.8871821751165383e+06, + "cpu_time": 1.9568567169810778e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5560352829135780e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 371, + "real_time": 1.8868794406806882e+06, + "cpu_time": 1.9567112398923112e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5566058198074699e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 371, + "real_time": 1.8873145729127638e+06, + "cpu_time": 1.9570511212937052e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5557858219908913e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 371, + "real_time": 1.8866478366340508e+06, + "cpu_time": 1.9565170754714750e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5570424271510170e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8869542134574284e+06, + "cpu_time": 1.9567460442048144e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5564649498882294e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time_median", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8868794406806887e+06, + "cpu_time": 1.9567112398923114e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.5566058198074699e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8469422957786486e+02, + "cpu_time": 2.1333979419330618e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3656363955229018e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5087500668933845e-04, + "cpu_time": 1.0902783977774875e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5086993604960258e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 187, + "real_time": 3.7477242581068194e+06, + "cpu_time": 3.8791317807484018e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5813127849432739e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 187, + "real_time": 3.7486371696294150e+06, + "cpu_time": 3.8797838074863059e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5804406221919998e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 187, + "real_time": 3.7486808309779768e+06, + "cpu_time": 3.8804396363637708e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5803989203579262e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 187, + "real_time": 3.7489007538393699e+06, + "cpu_time": 3.8806970213909638e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5801888823688736e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 187, + "real_time": 3.7485909284675186e+06, + "cpu_time": 3.8810409251335147e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5804847891170204e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7485067882042201e+06, + "cpu_time": 3.8802186342245913e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5805651997958183e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time_median", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7486371696294150e+06, + "cpu_time": 3.8804396363637708e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5804406221919998e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5330234666635971e+02, + "cpu_time": 7.6226362575001860e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3305100666595846e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2092877838525169e-04, + "cpu_time": 1.9644862766923609e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2094487392399758e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 94, + "real_time": 7.4779554616976930e+06, + "cpu_time": 7.7114002978719128e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5896904892645363e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 94, + "real_time": 7.4783486631163890e+06, + "cpu_time": 7.7114855425522961e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5895017482126480e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 94, + "real_time": 7.4793716070262045e+06, + "cpu_time": 7.7128112234035181e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5890108167353096e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 94, + "real_time": 7.4770856410899060e+06, + "cpu_time": 7.7115020744689358e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5901080833530640e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 94, + "real_time": 7.4781954655067082e+06, + "cpu_time": 7.7115905319151636e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5895752824082047e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4781913676873818e+06, + "cpu_time": 7.7117579340423662e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5895772839947525e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time_median", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4781954655067082e+06, + "cpu_time": 7.7115020744689358e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5895752824082047e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.2097600362140292e+02, + "cpu_time": 5.9266490867090329e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9406795622684169e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0978269520739590e-04, + "cpu_time": 7.6852115138971805e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0978115946518726e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 47, + "real_time": 1.4941837598985814e+07, + "cpu_time": 1.5365463659575054e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5930715244585472e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 47, + "real_time": 1.4938668169556780e+07, + "cpu_time": 1.5362947276595691e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5938338405165108e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 47, + "real_time": 1.4938920181482397e+07, + "cpu_time": 1.5362803148936439e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5937732143818581e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 47, + "real_time": 1.4939182180356473e+07, + "cpu_time": 1.5362689531914528e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5937101878704674e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 47, + "real_time": 1.4940292039450178e+07, + "cpu_time": 1.5364400191488806e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5934432244187744e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4939780033966329e+07, + "cpu_time": 1.5363660761702105e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5935663983292320e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time_median", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4939182180356473e+07, + "cpu_time": 1.5362947276595693e+07, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5937101878704674e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3065930132538626e+03, + "cpu_time": 1.2232176027102794e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1426325290749473e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.7457312643376181e-05, + "cpu_time": 7.9617587349980129e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7451633856996811e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 23, + "real_time": 2.9843611804687459e+07, + "cpu_time": 3.0646972304349102e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.5978950236557831e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 23, + "real_time": 2.9850422121260475e+07, + "cpu_time": 3.0653481043476976e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.5970741708045891e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 23, + "real_time": 2.9844042966547221e+07, + "cpu_time": 3.0647140434780337e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.5978430442670876e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 23, + "real_time": 2.9848202982026599e+07, + "cpu_time": 3.0651950782607462e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.5973416042720047e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 23, + "real_time": 2.9846303731850956e+07, + "cpu_time": 3.0648937739131369e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.5975705187712723e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9846516721274544e+07, + "cpu_time": 3.0649696460869052e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.5975448723541466e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time_median", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9846303731850952e+07, + "cpu_time": 3.0648937739131369e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.5975705187712723e+10, + "own_gpu": 5.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8591516504861993e+03, + "cpu_time": 2.9126872600641914e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4461864817795334e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.5795153491000224e-05, + "cpu_time": 9.5031520582360902e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.5792730988910008e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57084, + "real_time": 1.2261890215419478e+04, + "cpu_time": 3.3540234093616593e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3404311472706139e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57084, + "real_time": 1.2258462008974715e+04, + "cpu_time": 3.3527390302012354e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3413653335966778e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57084, + "real_time": 1.2259773622601197e+04, + "cpu_time": 3.3369967293811365e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3410078571507406e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57084, + "real_time": 1.2255505202027158e+04, + "cpu_time": 3.3371788907575705e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3421714833285612e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57084, + "real_time": 1.2253279289312737e+04, + "cpu_time": 3.3403986686286240e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3427786172902429e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2257782067667056e+04, + "cpu_time": 3.3442673456660457e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3415508877273679e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time_median", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2258462008974715e+04, + "cpu_time": 3.3403986686286240e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3413653335966778e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4192330747800037e+00, + "cpu_time": 8.4413407386074752e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.3214671699255588e+04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7894386242998065e-04, + "cpu_time": 2.5241225853390300e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7895631349385822e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57092, + "real_time": 1.2254569777059996e+04, + "cpu_time": 3.3441265133468609e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6848532009137166e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57092, + "real_time": 1.2250127466719588e+04, + "cpu_time": 3.3382765974216345e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6872773546687853e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57092, + "real_time": 1.2257251306693242e+04, + "cpu_time": 3.3404604568065020e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6833907497080064e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57092, + "real_time": 1.2263895922021316e+04, + "cpu_time": 3.3470145221747058e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6797696686990535e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57092, + "real_time": 1.2265241481948511e+04, + "cpu_time": 3.3446781493029783e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6790368636905003e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2258217190888532e+04, + "cpu_time": 3.3429112478105366e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6828655675360131e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time_median", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2257251306693244e+04, + "cpu_time": 3.3441265133468609e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6833907497080064e+08, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3495836306828499e+00, + "cpu_time": 3.4972143373203060e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4617167405782931e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.1798589728059812e-04, + "cpu_time": 1.0461582967872184e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1799885926100348e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57092, + "real_time": 1.2260011389988596e+04, + "cpu_time": 3.3758588698942724e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3363772250145714e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57092, + "real_time": 1.2260731841870811e+04, + "cpu_time": 3.3726189781406014e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3362986982594373e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57092, + "real_time": 1.2261822009944455e+04, + "cpu_time": 3.3762720503748431e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3361798912683954e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57092, + "real_time": 1.2261020251877153e+04, + "cpu_time": 3.3857088961675472e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3362672651561458e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57092, + "real_time": 1.2259284651565964e+04, + "cpu_time": 3.3875007409094229e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3364564463316510e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2260574029049396e+04, + "cpu_time": 3.3795919070973374e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3363159052060404e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time_median", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2260731841870813e+04, + "cpu_time": 3.3762720503748438e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3362986982594373e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6960466286221147e-01, + "cpu_time": 6.5868422102360000e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0568071006574473e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.9083137589226567e-05, + "cpu_time": 1.9490052027889085e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.9083628095746056e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57074, + "real_time": 1.2260231231472846e+04, + "cpu_time": 3.4166200879559277e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6727065241544805e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57074, + "real_time": 1.2260901204050464e+04, + "cpu_time": 3.4188575428392964e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6725604794185023e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57074, + "real_time": 1.2260004825392425e+04, + "cpu_time": 3.4024525931247517e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6727558811504087e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57074, + "real_time": 1.2259959234488568e+04, + "cpu_time": 3.4024580527035730e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6727658202826757e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57074, + "real_time": 1.2262709748558957e+04, + "cpu_time": 3.4071553544522256e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6721663214650178e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2260761248792651e+04, + "cpu_time": 3.4095087262151545e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6725910052942171e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time_median", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2260231231472846e+04, + "cpu_time": 3.4071553544522249e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6727065241544805e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1524496196347007e+00, + "cpu_time": 7.7944083785239314e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5118574991428156e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.3994948294763130e-05, + "cpu_time": 2.2860796098258967e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.3985854706799519e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57057, + "real_time": 1.2275330744436977e+04, + "cpu_time": 3.4871695217063861e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3388378174413004e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57057, + "real_time": 1.2271480318828813e+04, + "cpu_time": 3.4830350000875645e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3405129859878826e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57057, + "real_time": 1.2265687225050648e+04, + "cpu_time": 3.4753257794836340e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3430353144953432e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57057, + "real_time": 1.2269444882980295e+04, + "cpu_time": 3.4791396182764016e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3413989487746944e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57057, + "real_time": 1.2273103994496210e+04, + "cpu_time": 3.4768654117811544e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3398064604837675e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2271009433158590e+04, + "cpu_time": 3.4803070662670281e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3407183054365978e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time_median", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2271480318828813e+04, + "cpu_time": 3.4791396182764016e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3405129859878826e+09, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6761517913194481e+00, + "cpu_time": 4.8103029015986756e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6001295819526617e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9958022698489581e-04, + "cpu_time": 1.3821489914561474e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9960943274686585e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57092, + "real_time": 1.2262225427540754e+04, + "cpu_time": 3.6185585896447075e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0689087455986126e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57092, + "real_time": 1.2260647363350616e+04, + "cpu_time": 3.6201989000211426e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0690463245178953e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57092, + "real_time": 1.2260233032232083e+04, + "cpu_time": 3.5951123169620703e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0690824526370131e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57092, + "real_time": 1.2259557543402589e+04, + "cpu_time": 3.6142917764309350e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0691413579647142e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57092, + "real_time": 1.2260022705512840e+04, + "cpu_time": 3.6178363816296129e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0691007932722847e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2260537214407776e+04, + "cpu_time": 3.6131995929376935e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0690559347981041e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time_median", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2260233032232081e+04, + "cpu_time": 3.6178363816296129e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0690824526370131e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0221543268822102e+00, + "cpu_time": 1.0338657694858223e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9120163159635209e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.3369456738081744e-05, + "cpu_time": 2.8613580370887926e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.3363423988161984e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 44705, + "real_time": 1.5657522508725173e+04, + "cpu_time": 4.1370846818029087e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6742367756707355e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 44705, + "real_time": 1.5658348501274942e+04, + "cpu_time": 4.1296303254668564e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6741484581126520e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 44705, + "real_time": 1.5654268575138494e+04, + "cpu_time": 4.1216390716920876e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6745847865184004e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 44705, + "real_time": 1.5658413058940181e+04, + "cpu_time": 4.1289556470192285e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6741415558093784e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 44705, + "real_time": 1.5654712485068107e+04, + "cpu_time": 4.1252827692652514e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6745373014677858e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5656653025829379e+04, + "cpu_time": 4.1285184990492671e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6743297755157906e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time_median", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5657522508725171e+04, + "cpu_time": 4.1289556470192285e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6742367756707355e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0111972625057866e+00, + "cpu_time": 5.7576733862937807e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1508676253456417e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2845639864330122e-04, + "cpu_time": 1.3946100490090289e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2846140926347976e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 36875, + "real_time": 1.8984367859735295e+04, + "cpu_time": 4.4748152949151990e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7616826847944901e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 36875, + "real_time": 1.8979591554103783e+04, + "cpu_time": 4.4703122630510159e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7623776755440136e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 36875, + "real_time": 1.8982403192637583e+04, + "cpu_time": 4.4701080542373827e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7619685172599625e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 36875, + "real_time": 1.8977204543860011e+04, + "cpu_time": 4.4598541396611588e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7627251357715431e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 36875, + "real_time": 1.8981127369551035e+04, + "cpu_time": 4.4658698413556936e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7621541639357384e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8980938903977541e+04, + "cpu_time": 4.4681919186440908e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7621816354611500e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time_median", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8981127369551039e+04, + "cpu_time": 4.4701080542373827e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.7621541639357384e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7248025094715791e+00, + "cpu_time": 5.6335413048184407e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9653187534724115e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4355467467947985e-04, + "cpu_time": 1.2608100563701807e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4355749464717573e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27325, + "real_time": 2.5608575346724476e+04, + "cpu_time": 5.1476323915828805e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.0946284039737511e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27325, + "real_time": 2.5614729927826545e+04, + "cpu_time": 5.1434225178408487e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.0936445668352722e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27325, + "real_time": 2.5610038652371277e+04, + "cpu_time": 5.1375129515097498e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.0943944452145935e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27325, + "real_time": 2.5612539984208612e+04, + "cpu_time": 5.1329873559009284e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.0939945848654549e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27325, + "real_time": 2.5610074945819764e+04, + "cpu_time": 5.1348215407138319e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.0943886428225983e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5611191771390142e+04, + "cpu_time": 5.1392753515096483e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.0942101287423340e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time_median", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5610074945819764e+04, + "cpu_time": 5.1375129515097491e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.0943886428225983e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4371545532109087e+00, + "cpu_time": 6.1139283707053309e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8958742451983741e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.5159747932285443e-05, + "cpu_time": 1.1896479469443841e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.5155698478893532e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18064, + "real_time": 3.8702415077767546e+04, + "cpu_time": 6.4854706875553689e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 5.4186592639917732e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18064, + "real_time": 3.8686304345614808e+04, + "cpu_time": 6.4784892548714051e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 5.4209158395294426e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18064, + "real_time": 3.8690288854792474e+04, + "cpu_time": 6.4699785595656605e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 5.4203575679436432e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18064, + "real_time": 3.8745940424951878e+04, + "cpu_time": 6.4719088795395961e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 5.4125721998206085e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18064, + "real_time": 3.8686507697231231e+04, + "cpu_time": 6.4585273306026604e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 5.4208873450474106e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8702291280071593e+04, + "cpu_time": 6.4728749424269387e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 5.4186784432665764e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time_median", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8690288854792481e+04, + "cpu_time": 6.4719088795395961e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 5.4203575679436432e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5268680949928498e+01, + "cpu_time": 1.0066647602876718e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5352236722737640e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.5289883658489582e-04, + "cpu_time": 1.5552050197809525e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5241436805809109e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 11347, + "real_time": 6.1801748240772526e+04, + "cpu_time": 8.8604384066265338e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.7867076893350533e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 11347, + "real_time": 6.1697005059267947e+04, + "cpu_time": 8.8363602009333976e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.7982295023410439e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 11347, + "real_time": 6.1804976276084235e+04, + "cpu_time": 8.8339598307919689e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.7863532238309555e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 11347, + "real_time": 6.1711713765708082e+04, + "cpu_time": 8.8158485679036443e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.7966091752433037e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 11347, + "real_time": 6.1801704443957780e+04, + "cpu_time": 8.8278549572574586e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.7867124988493225e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1763429557158110e+04, + "cpu_time": 8.8348923927026015e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.7909224179199371e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time_median", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1801704443957773e+04, + "cpu_time": 8.8339598307919689e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.7867124988493225e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4189840681876738e+01, + "cpu_time": 1.6338724862714452e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9602345116600908e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.7737745572770540e-04, + "cpu_time": 1.8493405619982226e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7767671972399197e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 6085, + "real_time": 1.1494432716210844e+05, + "cpu_time": 1.4311416088742725e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2979747736218124e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 6085, + "real_time": 1.1500420274991418e+05, + "cpu_time": 1.4316839244044069e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2941751687472656e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 6085, + "real_time": 1.1494913915165141e+05, + "cpu_time": 1.4304901495481000e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2976692665205444e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 6085, + "real_time": 1.1502234437152815e+05, + "cpu_time": 1.4313837715695403e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2930247125761597e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 6085, + "real_time": 1.1495129568331079e+05, + "cpu_time": 1.4308933541495795e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2975323593659149e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1497426182370260e+05, + "cpu_time": 1.4311185617091798e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2960752561663406e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time_median", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1495129568331079e+05, + "cpu_time": 1.4311416088742722e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2975323593659149e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6273498153800688e+01, + "cpu_time": 4.5719872759802051e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3015393537619989e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1549233348783024e-04, + "cpu_time": 3.1946949737832322e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1544896029092259e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3146, + "real_time": 2.2249025823723839e+05, + "cpu_time": 2.5632302256835319e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.5406519516511505e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3146, + "real_time": 2.2250480420866155e+05, + "cpu_time": 2.5642352542909456e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.5401589910241165e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3146, + "real_time": 2.2251365121805691e+05, + "cpu_time": 2.5651400381437072e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.5398591988222855e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3146, + "real_time": 2.2251497365226282e+05, + "cpu_time": 2.5644752479338963e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.5398143884998672e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3146, + "real_time": 2.2246131996372691e+05, + "cpu_time": 2.5641163127782015e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.5416328567750946e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2249700145598935e+05, + "cpu_time": 2.5642394157660566e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.5404234773545029e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time_median", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2250480420866158e+05, + "cpu_time": 2.5642352542909453e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.5401589910241165e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2247119085073525e+01, + "cpu_time": 6.8927022381456851e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5401451848939881e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.9988399571641323e-05, + "cpu_time": 2.6880104079854482e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9996309325844218e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1596, + "real_time": 4.3857730154451658e+05, + "cpu_time": 4.8439061779446172e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6507452350664230e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1596, + "real_time": 4.3864095836885413e+05, + "cpu_time": 4.8440543859651324e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6496349371423737e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1596, + "real_time": 4.3853429691497830e+05, + "cpu_time": 4.8411639348370518e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6514955012755661e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1596, + "real_time": 4.3860697687500244e+05, + "cpu_time": 4.8420137218043057e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6502275999049133e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1596, + "real_time": 4.3851775357562152e+05, + "cpu_time": 4.8415920614033862e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6517841584294266e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3857545745579462e+05, + "cpu_time": 4.8425460563908983e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6507774863637405e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time_median", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3857730154451664e+05, + "cpu_time": 4.8420137218043057e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.6507452350664230e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0771970341922426e+01, + "cpu_time": 1.3443153744804277e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8568731077463217e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1576564415267102e-04, + "cpu_time": 2.7760507774753779e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1576435366905193e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 802, + "real_time": 8.7244182709253801e+05, + "cpu_time": 9.4279990523694071e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.6920732037394516e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 802, + "real_time": 8.7239293831846502e+05, + "cpu_time": 9.4288081670821935e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.6925042664091431e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 802, + "real_time": 8.7232790773262992e+05, + "cpu_time": 9.4261525685781357e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.6930777297301590e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 802, + "real_time": 8.7247872624971310e+05, + "cpu_time": 9.4283343391519284e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.6917478880502472e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 802, + "real_time": 8.7242866025835811e+05, + "cpu_time": 9.4271304613469914e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.6921892937500015e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7241401193034090e+05, + "cpu_time": 9.4276849177057331e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.6923184763358017e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time_median", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.7242866025835811e+05, + "cpu_time": 9.4279990523694060e+05, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.6921892937500015e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.7083775766897006e+01, + "cpu_time": 1.0532845769756265e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0333919000848718e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.5431979526086453e-05, + "cpu_time": 1.1172250517170954e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5434002967626779e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 402, + "real_time": 1.7401005575243400e+06, + "cpu_time": 1.8575510995025134e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.7132167689752945e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 402, + "real_time": 1.7401097114753926e+06, + "cpu_time": 1.8575667014924400e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.7131761931378677e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 402, + "real_time": 1.7400718866314609e+06, + "cpu_time": 1.8576384800993609e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.7133438584440903e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 402, + "real_time": 1.7400886069992511e+06, + "cpu_time": 1.8575105522389058e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.7132697415596466e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 402, + "real_time": 1.7401272134475447e+06, + "cpu_time": 1.8575589502489013e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.7130986150195007e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7400995952155977e+06, + "cpu_time": 1.8575651567164245e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.7132210354272797e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time_median", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7401005575243402e+06, + "cpu_time": 1.8575589502489013e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.7132167689752945e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0943604851720256e+01, + "cpu_time": 4.6358590415079554e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.2835356574960169e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2035865596029489e-05, + "cpu_time": 2.4956642972904692e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2035874007572439e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 201, + "real_time": 3.4771777645439208e+06, + "cpu_time": 3.6997205223877388e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.7199232877071213e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 201, + "real_time": 3.4771546946180547e+06, + "cpu_time": 3.6999876865672437e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.7199745071878677e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 201, + "real_time": 3.4772916129368604e+06, + "cpu_time": 3.7001623383082286e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.7196705332770203e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 201, + "real_time": 3.4772263496965910e+06, + "cpu_time": 3.7000458059704066e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.7198154219503891e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 201, + "real_time": 3.4772016742818095e+06, + "cpu_time": 3.7000759601988057e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.7198702044063461e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4772104192154473e+06, + "cpu_time": 3.6999984626864851e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.7198507909057495e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time_median", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4772016742818095e+06, + "cpu_time": 3.7000458059704066e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.7198702044063461e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2664567478282869e+01, + "cpu_time": 1.6768228151973415e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1692106033046399e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5145637200225982e-05, + "cpu_time": 4.5319554375702053e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5145507795073065e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 101, + "real_time": 6.9536661664669467e+06, + "cpu_time": 7.3782772673257114e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.7206886144316589e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 101, + "real_time": 6.9533002028002003e+06, + "cpu_time": 7.3779065742572509e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.7210949670171555e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 101, + "real_time": 6.9539053826638963e+06, + "cpu_time": 7.3782628910892904e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.7204230206873459e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 101, + "real_time": 6.9536820542768091e+06, + "cpu_time": 7.3776275346528906e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.7206709741611160e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 101, + "real_time": 6.9536803991045104e+06, + "cpu_time": 7.3779166930690547e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.7206728118988297e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9536468410624731e+06, + "cpu_time": 7.3779981920788391e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.7207100776392212e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time_median", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.9536803991045104e+06, + "cpu_time": 7.3779166930690538e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.7206728118988297e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1779657800046374e+02, + "cpu_time": 2.7403076125755689e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4182641322072325e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1321202094178517e-05, + "cpu_time": 3.7141614042649344e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1321783979572389e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 50, + "real_time": 1.3912737444043159e+07, + "cpu_time": 1.4716493779999383e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.7176891199059494e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 50, + "real_time": 1.3911940604448318e+07, + "cpu_time": 1.4716060180001022e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.7181311689662689e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 50, + "real_time": 1.3913163095712662e+07, + "cpu_time": 1.4717757559999429e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.7174530091642014e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 50, + "real_time": 1.3912791814655066e+07, + "cpu_time": 1.4716280580000784e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.7176589594977768e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 50, + "real_time": 1.3913214355707169e+07, + "cpu_time": 1.4716734219998669e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.7174245760078690e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3912769462913275e+07, + "cpu_time": 1.4716665263999859e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.7176713667084137e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time_median", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3912791814655066e+07, + "cpu_time": 1.4716493779999381e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.7176589594977768e+10, + "own_gpu": 6.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1024690758509394e+02, + "cpu_time": 6.5980694474248867e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8305159537582542e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6674718785877906e-05, + "cpu_time": 4.4833998253430346e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6675777177662710e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57207, + "real_time": 1.2260957072471429e+04, + "cpu_time": 3.1164959934971252e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3406853769975501e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57207, + "real_time": 1.2255022700369378e+04, + "cpu_time": 3.1172651913227590e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3423030704598719e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57207, + "real_time": 1.2260906625475851e+04, + "cpu_time": 3.1179390476689292e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3406991221100122e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57207, + "real_time": 1.2258261285819086e+04, + "cpu_time": 3.1150892600555366e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3414200468531692e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57207, + "real_time": 1.2244752590520920e+04, + "cpu_time": 3.1150114846086424e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3451063790140224e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2255980054931333e+04, + "cpu_time": 3.1163601954305981e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3420427990869254e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time_median", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2258261285819086e+04, + "cpu_time": 3.1164959934971248e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.3414200468531692e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.7308673302959265e+00, + "cpu_time": 1.3004355684168184e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8363786406947777e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.4919046050402854e-04, + "cpu_time": 4.1729308772573787e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4947789453698554e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56959, + "real_time": 1.2243345222691767e+04, + "cpu_time": 3.1130226074896156e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6909817954140341e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56959, + "real_time": 1.2242921934574038e+04, + "cpu_time": 3.1201710054599997e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6912131301481020e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56959, + "real_time": 1.2252241255526656e+04, + "cpu_time": 3.1193639828649400e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6861236480344427e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56959, + "real_time": 1.2251778468554932e+04, + "cpu_time": 3.1197908442915712e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6863762032796752e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56959, + "real_time": 1.2243329173708564e+04, + "cpu_time": 3.1213481802701419e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6909905661865032e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2246723211011195e+04, + "cpu_time": 3.1187393240752543e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6891370686125517e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time_median", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2243345222691767e+04, + "cpu_time": 3.1197908442915716e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.6909817954140341e+08, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8317812678714889e+00, + "cpu_time": 3.2799612846368333e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6387299354045308e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.9453665969417593e-04, + "cpu_time": 1.0516945931700733e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9447987211777246e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57266, + "real_time": 1.2230684025568908e+04, + "cpu_time": 3.1231239828171314e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3395816591899815e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57266, + "real_time": 1.2218672308094327e+04, + "cpu_time": 3.1314887612195624e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3408985515673687e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57266, + "real_time": 1.2223481086739876e+04, + "cpu_time": 3.1361764886669731e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3403710353651619e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57266, + "real_time": 1.2223416525988636e+04, + "cpu_time": 3.1349175025321561e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3403781148392842e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57266, + "real_time": 1.2221156569128731e+04, + "cpu_time": 3.1344852844621564e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3406259798182135e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2223482103104096e+04, + "cpu_time": 3.1320384039395962e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3403710681560020e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time_median", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2223416525988634e+04, + "cpu_time": 3.1344852844621561e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.3403781148392842e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.4836989446221720e+00, + "cpu_time": 5.2718400315464308e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9153709833541559e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6681028423836431e-04, + "cpu_time": 1.6831977618522529e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6671718005047759e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57234, + "real_time": 1.2230801818380267e+04, + "cpu_time": 3.1589970507041045e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6791375158051157e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57234, + "real_time": 1.2233723384759289e+04, + "cpu_time": 3.1713429167976868e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6784977042085333e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57234, + "real_time": 1.2236110629675159e+04, + "cpu_time": 3.1736003756508449e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6779751337431245e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57234, + "real_time": 1.2235084474805937e+04, + "cpu_time": 3.1682911888038387e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6781997351530132e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57234, + "real_time": 1.2235948664359361e+04, + "cpu_time": 3.1720462137890328e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6780105816761074e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2234333794396003e+04, + "cpu_time": 3.1688555491491017e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6783641341171789e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time_median", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2235084474805935e+04, + "cpu_time": 3.1713429167976865e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.6781997351530132e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1895161642187189e+00, + "cpu_time": 5.8392895839434118e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7940271238281496e+05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7896488693333165e-04, + "cpu_time": 1.8427124535579963e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7899086471333438e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 57081, + "real_time": 1.2258094574876677e+04, + "cpu_time": 3.2292769678177454e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3463447846387110e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 57081, + "real_time": 1.2263402816699901e+04, + "cpu_time": 3.2445474588741879e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3440306071293058e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 57081, + "real_time": 1.2262501097420194e+04, + "cpu_time": 3.2408643892013981e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3444235787907553e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 57081, + "real_time": 1.2264002663501244e+04, + "cpu_time": 3.2412546363939149e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3437692243039808e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 57081, + "real_time": 1.2263764633484740e+04, + "cpu_time": 3.2417731189011516e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3438729426575747e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2262353157196552e+04, + "cpu_time": 3.2395433142376794e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3444882275040665e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time_median", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2263402816699900e+04, + "cpu_time": 3.2412546363939146e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 5.3440306071293058e+09, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4480773578193307e+00, + "cpu_time": 5.9178801901024016e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0672325488289795e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9964172670908589e-04, + "cpu_time": 1.8267637182356924e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9968844600250688e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 56956, + "real_time": 1.2284831303140652e+04, + "cpu_time": 3.3656538099584941e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0669417981058565e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 56956, + "real_time": 1.2285809616455834e+04, + "cpu_time": 3.3805630153100981e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0668568380258783e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 56956, + "real_time": 1.2289175769498994e+04, + "cpu_time": 3.3802966974506940e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0665646131070314e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 56956, + "real_time": 1.2288946945927140e+04, + "cpu_time": 3.3830181596320726e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0665844728334555e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 56956, + "real_time": 1.2289420256560546e+04, + "cpu_time": 3.3808280497226304e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0665433947547600e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2287636778316633e+04, + "cpu_time": 3.3780719464147973e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0666982233653965e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time_median", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2288946945927139e+04, + "cpu_time": 3.3805630153100981e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.0665844728334555e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1491299956036198e+00, + "cpu_time": 7.0254214303027965e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8658111276761107e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7490181671027903e-04, + "cpu_time": 2.0797133814035515e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7491461847471166e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 43266, + "real_time": 1.6182874302440559e+04, + "cpu_time": 3.9329253871400491e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6198852879952587e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 43266, + "real_time": 1.6182919232563512e+04, + "cpu_time": 3.9380457703508830e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6198807905591589e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 43266, + "real_time": 1.6185555949103271e+04, + "cpu_time": 3.9367459853003798e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6196169030234859e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 43266, + "real_time": 1.6184863558497318e+04, + "cpu_time": 3.9394522581239566e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6196861904490393e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 43266, + "real_time": 1.6185671191116813e+04, + "cpu_time": 3.9369702537787962e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6196053713476681e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6184376846744297e+04, + "cpu_time": 3.9368279309388134e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6197349086749222e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time_median", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6184863558497314e+04, + "cpu_time": 3.9369702537787962e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.6196861904490393e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3860714890316514e+00, + "cpu_time": 2.4303829427154774e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3872110652384518e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.5642561474986802e-05, + "cpu_time": 6.1734548355937542e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.5644327217304083e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 35168, + "real_time": 1.9902199883447687e+04, + "cpu_time": 4.3178608308691051e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6343218491943756e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 35168, + "real_time": 1.9907245755604959e+04, + "cpu_time": 4.3340142487487632e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6336541299410282e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 35168, + "real_time": 1.9902290423065951e+04, + "cpu_time": 4.3215523544130148e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6343098651217113e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 35168, + "real_time": 1.9904620585680630e+04, + "cpu_time": 4.3253008530479521e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6340014759044060e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 35168, + "real_time": 1.9904716432425990e+04, + "cpu_time": 4.3232532813923412e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6339887924547523e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9904214616045043e+04, + "cpu_time": 4.3243963136942351e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6340552225232544e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time_median", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9904620585680626e+04, + "cpu_time": 4.3232532813923419e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.6340014759044060e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0836328942975122e+00, + "cpu_time": 6.0297186735590813e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7573064672321067e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0468299978125583e-04, + "cpu_time": 1.3943492307734457e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0467914429640491e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 25606, + "real_time": 2.7337470635266047e+04, + "cpu_time": 5.1089833671796659e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8356730730139671e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 25606, + "real_time": 2.7338390026290272e+04, + "cpu_time": 5.1175977036632234e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8355440791927574e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 25606, + "real_time": 2.7344070128935098e+04, + "cpu_time": 5.1157565570568542e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8347473329890717e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 25606, + "real_time": 2.7345274184910333e+04, + "cpu_time": 5.1233575490118405e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8345784829563904e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 25606, + "real_time": 2.7342227246594975e+04, + "cpu_time": 5.1192590252282767e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8350057972346893e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7341486444399350e+04, + "cpu_time": 5.1169908404279726e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8351097530773750e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time_median", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7342227246594975e+04, + "cpu_time": 5.1175977036632234e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.8350057972346893e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4382306371433473e+00, + "cpu_time": 5.2841983008146599e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8227886473781951e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2575141604444980e-04, + "cpu_time": 1.0326769121933200e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2575360179740058e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 16671, + "real_time": 4.1931743350214230e+04, + "cpu_time": 6.6641947573635873e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 5.0013470283946243e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 16671, + "real_time": 4.1991267478931040e+04, + "cpu_time": 6.6756989082834669e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9942574394836685e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 16671, + "real_time": 4.1919234626099111e+04, + "cpu_time": 6.6718153140184804e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 5.0028394332712921e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 16671, + "real_time": 4.1996690055792395e+04, + "cpu_time": 6.6876581188885481e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9936125852155113e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 16671, + "real_time": 4.2009340165280628e+04, + "cpu_time": 6.6840008397818703e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9921088780471466e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1969655135263485e+04, + "cpu_time": 6.6766735876671912e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9968330728824493e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time_median", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1991267478931040e+04, + "cpu_time": 6.6756989082834654e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9942574394836685e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1086534057036204e+01, + "cpu_time": 9.4157126419410744e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8932505683867857e+07, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.7895810495985552e-04, + "cpu_time": 1.4102400721421064e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.7927036925491849e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10264, + "real_time": 6.8199322885535250e+04, + "cpu_time": 9.4195721063910707e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.1500669252093002e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10264, + "real_time": 6.8191607933330321e+04, + "cpu_time": 9.4266060502729466e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.1507627215664040e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10264, + "real_time": 6.8186806416613923e+04, + "cpu_time": 9.4134987431801273e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.1511958403994781e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10264, + "real_time": 6.8185746505807867e+04, + "cpu_time": 9.4205118959468862e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.1512914574349350e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10264, + "real_time": 6.8194756660698360e+04, + "cpu_time": 9.4216452650035848e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.1504787250267273e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8191648080397150e+04, + "cpu_time": 9.4203668121589231e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.1507591339273697e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time_median", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8191607933330321e+04, + "cpu_time": 9.4205118959468862e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.1507627215664040e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.6311145290639590e+00, + "cpu_time": 4.7006784609475588e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0790643225224074e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.2577774369449781e-05, + "cpu_time": 4.9899102175940380e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2576218836248497e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5522, + "real_time": 1.2677683264737017e+05, + "cpu_time": 1.5441367928287905e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 6.6168303978163879e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5522, + "real_time": 1.2674266900372636e+05, + "cpu_time": 1.5446572546178967e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 6.6186139726577522e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5522, + "real_time": 1.2678244130124550e+05, + "cpu_time": 1.5435044223108396e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 6.6165376797469749e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5522, + "real_time": 1.2675655222392807e+05, + "cpu_time": 1.5430065483519694e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 6.6178890580588593e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5522, + "real_time": 1.2677820796411020e+05, + "cpu_time": 1.5436003839189018e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 6.6167586170446114e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2676734062807607e+05, + "cpu_time": 1.5437810804056798e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 6.6173259450649170e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time_median", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2677683264737017e+05, + "cpu_time": 1.5436003839189018e+05, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 6.6168303978163879e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7038973440097781e+01, + "cpu_time": 6.3310808737298984e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8950853515410405e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3441138195119663e-04, + "cpu_time": 4.1010224533041931e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3442114572238708e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2865, + "real_time": 2.4434878964450074e+05, + "cpu_time": 2.7799670924956398e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.8660933513969566e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2865, + "real_time": 2.4431075356912069e+05, + "cpu_time": 2.7794436300173466e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.8671623147580238e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2865, + "real_time": 2.4436649677049476e+05, + "cpu_time": 2.7805136684119818e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.8655958250107010e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2865, + "real_time": 2.4434376223460588e+05, + "cpu_time": 2.7807827574170375e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.8662346223069977e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2865, + "real_time": 2.4436158002118531e+05, + "cpu_time": 2.7809285479931737e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.8657339662583099e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4434627644798151e+05, + "cpu_time": 2.7803271392670355e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.8661640159461975e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time_median", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4434878964450076e+05, + "cpu_time": 2.7805136684119818e+05, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 6.8660933513969566e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1895051306403349e+01, + "cpu_time": 6.1528362354925413e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1529727547194613e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.9606650138843229e-05, + "cpu_time": 2.2129900286174897e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9612959149091133e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1460, + "real_time": 4.7949376222613739e+05, + "cpu_time": 5.2465781095891120e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 6.9978870724443665e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1460, + "real_time": 4.7959699522710266e+05, + "cpu_time": 5.2479366849314678e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 6.9963807809327560e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1460, + "real_time": 4.7952828122293320e+05, + "cpu_time": 5.2477668424654589e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 6.9973833273037979e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1460, + "real_time": 4.7958504232380236e+05, + "cpu_time": 5.2467276301365311e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 6.9965551547258209e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1460, + "real_time": 4.7948246796644124e+05, + "cpu_time": 5.2457814726030594e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 6.9980519084064728e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7953730979328335e+05, + "cpu_time": 5.2469581479451270e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 6.9972516487626419e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time_median", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7952828122293326e+05, + "cpu_time": 5.2467276301365311e+05, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 6.9973833273037979e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2024698425985640e+01, + "cpu_time": 8.9353985661167002e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5911672704953616e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0848936540185412e-04, + "cpu_time": 1.7029673792263967e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0848784139180910e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 738, + "real_time": 9.4906721269802703e+05, + "cpu_time": 1.0180919525746652e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.0710338637894348e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 738, + "real_time": 9.4905680780258623e+05, + "cpu_time": 1.0181658983740603e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.0711113864070557e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 738, + "real_time": 9.4918190292776795e+05, + "cpu_time": 1.0181512899728026e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.0701794664438446e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 738, + "real_time": 9.4918883705340070e+05, + "cpu_time": 1.0181699308942569e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.0701278165394714e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 738, + "real_time": 9.4921506047223613e+05, + "cpu_time": 1.0181914607045505e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.0699324941824280e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.4914196419080370e+05, + "cpu_time": 1.0181541065040672e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.0704770054724472e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time_median", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.4918190292776807e+05, + "cpu_time": 1.0181658983740604e+06, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.0701794664438446e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4119211705820817e+01, + "cpu_time": 3.7602655677559220e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5215153936143294e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.8090754072823626e-05, + "cpu_time": 3.6932184860179618e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8092544383367009e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 371, + "real_time": 1.8865249221366290e+06, + "cpu_time": 2.0029889514822485e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.1145483648309555e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 371, + "real_time": 1.8864783726386586e+06, + "cpu_time": 2.0030472264151266e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.1147239187410736e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 371, + "real_time": 1.8866043035794462e+06, + "cpu_time": 2.0031703692723266e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.1142490105290909e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 371, + "real_time": 1.8864947489169086e+06, + "cpu_time": 2.0031719083555427e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.1146621572659180e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 371, + "real_time": 1.8864075720405604e+06, + "cpu_time": 2.0030024285714694e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.1149909483672348e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8865019838624406e+06, + "cpu_time": 2.0030761768193431e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.1146348799468552e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time_median", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8864947489169086e+06, + "cpu_time": 2.0030472264151264e+06, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.1146621572659180e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1622127163071326e+01, + "cpu_time": 8.9333547292740917e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7010957790644891e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7965572141319226e-05, + "cpu_time": 4.4598177706148166e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7965346425263997e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 186, + "real_time": 3.7616185746305892e+06, + "cpu_time": 3.9970415053760437e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.1361689303217499e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 186, + "real_time": 3.7612530104916100e+06, + "cpu_time": 3.9990545268816883e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.1368625096803696e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 186, + "real_time": 3.7613493272714238e+06, + "cpu_time": 4.0001872473115041e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.1366797562174255e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 186, + "real_time": 3.7620891003258605e+06, + "cpu_time": 4.0002375537628671e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.1352764073755966e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 186, + "real_time": 3.7622843751363377e+06, + "cpu_time": 3.9977109892473188e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.1349060632949219e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7617188775711642e+06, + "cpu_time": 3.9988463645158848e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.1359787333780136e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time_median", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7616185746305897e+06, + "cpu_time": 3.9990545268816883e+06, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 7.1361689303217499e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5291075811271349e+02, + "cpu_time": 1.4425064344960381e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.5914947879772354e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2039994822929062e-04, + "cpu_time": 3.6073064654252435e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2039686648435698e-04, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 93, + "real_time": 7.5121331779706860e+06, + "cpu_time": 7.9356560107518369e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.1467171744821136e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 93, + "real_time": 7.5120281938823964e+06, + "cpu_time": 7.9357932580644768e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.1468170531789795e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 93, + "real_time": 7.5127008541296888e+06, + "cpu_time": 7.9361916344080986e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.1461771528528931e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 93, + "real_time": 7.5124531623817263e+06, + "cpu_time": 7.9363012043031445e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.1464127681801346e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 93, + "real_time": 7.5116514238298582e+06, + "cpu_time": 7.9350650107527897e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.1471755238380493e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5121933624388725e+06, + "cpu_time": 7.9358014236560687e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.1466599345064346e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time_median", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5121331779706869e+06, + "cpu_time": 7.9357932580644768e+06, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 7.1467171744821136e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0300250301165721e+02, + "cpu_time": 4.9131473346485274e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8339415033200495e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3646449654328435e-05, + "cpu_time": 6.1911167787071615e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3646620077842426e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 47, + "real_time": 1.5014225220743647e+07, + "cpu_time": 1.5817491446810851e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.1514967187019333e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 47, + "real_time": 1.5012614765858397e+07, + "cpu_time": 1.5816773510637853e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.1522638843827362e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 47, + "real_time": 1.5012999283189470e+07, + "cpu_time": 1.5816390446809934e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.1520806985070786e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 47, + "real_time": 1.5013213645904623e+07, + "cpu_time": 1.5816717872338176e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.1519785791691605e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 47, + "real_time": 1.5011160911873300e+07, + "cpu_time": 1.5814757382979728e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.1529565921227844e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5012842765513888e+07, + "cpu_time": 1.5816426131915310e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.1521552945767380e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time_median", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5012999283189470e+07, + "cpu_time": 1.5816717872338172e+07, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.1520806985070786e+10, + "own_gpu": 7.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1128620680396107e+03, + "cpu_time": 1.0157111569292431e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3018502358327704e+06, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_implicit_mapped_GPUWrGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.4127337868080143e-05, + "cpu_time": 6.4218752609331993e-05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.4129405996721046e-05, + "own_gpu": 0.0000000000000000e+00, + "wr_gpu": NaN + } + ] +} diff --git a/results/implicit_mapped_GPUWrHost.json b/results/implicit_mapped_GPUWrHost.json new file mode 100644 index 0000000..639133b --- /dev/null +++ b/results/implicit_mapped_GPUWrHost.json @@ -0,0 +1,1200 @@ +{ + "context": { + "date": "2023-10-10T13:17:42-04:00", + "host_name": "frontier00741", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1795, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [2.33,4.27,7.21], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52613, + "real_time": 1.3306196793273351e+04, + "cpu_time": 2.0417493072054429e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0782650096311820e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52613, + "real_time": 1.3300892996205879e+04, + "cpu_time": 2.0429356261760397e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0794924830749309e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52613, + "real_time": 1.3302283200872382e+04, + "cpu_time": 2.0441628190751337e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0791706492396575e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52613, + "real_time": 1.3304591109098894e+04, + "cpu_time": 2.0434165415391642e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0786365145778745e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52613, + "real_time": 1.3293467127319691e+04, + "cpu_time": 2.0402218140003402e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0812127195787936e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3301486245354041e+04, + "cpu_time": 2.0424972215992242e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0793554752204877e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3302283200872382e+04, + "cpu_time": 2.0429356261760393e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 3.0791706492396575e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9275444455396196e+00, + "cpu_time": 1.5450257565623494e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1410858880908132e+05, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7045066653816362e-04, + "cpu_time": 7.5643958788479187e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7055997505747835e-04, + "own_numa": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52606, + "real_time": 1.3301542443827195e+04, + "cpu_time": 2.0655706839524035e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1586842537961721e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52606, + "real_time": 1.3305887178059214e+04, + "cpu_time": 2.0655983195833174e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1566732758024776e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52606, + "real_time": 1.3302569747431944e+04, + "cpu_time": 2.0632565125651076e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1582086435453296e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52606, + "real_time": 1.3294035660672711e+04, + "cpu_time": 2.0610971391096064e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1621618965820229e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52606, + "real_time": 1.3305873130043401e+04, + "cpu_time": 2.0670776375318404e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1566797758677268e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3301981632006893e+04, + "cpu_time": 2.0645200585484552e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1584815691187465e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3302569747431942e+04, + "cpu_time": 2.0655706839524035e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 6.1582086435453296e+08, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8495946542418906e+00, + "cpu_time": 2.3516269256781630e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2459276088066597e+05, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6457685692280000e-04, + "cpu_time": 1.1390671240712334e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6468853297681346e-04, + "own_numa": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52631, + "real_time": 1.3304213196133955e+04, + "cpu_time": 2.0817869943569371e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2314895859276364e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52631, + "real_time": 1.3297664440957693e+04, + "cpu_time": 2.0802262373886118e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2320960626391044e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52631, + "real_time": 1.3294788606614275e+04, + "cpu_time": 2.0806361535976899e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2323625809175196e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52631, + "real_time": 1.3297339042672187e+04, + "cpu_time": 2.0788243924683138e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2321262131786275e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52631, + "real_time": 1.3297865139036545e+04, + "cpu_time": 2.0793954912504021e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2320774672247167e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3298374085082931e+04, + "cpu_time": 2.0801738538123911e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2320303819775207e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3297664440957693e+04, + "cpu_time": 2.0802262373886111e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.2320960626391044e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4923060864635134e+00, + "cpu_time": 1.1451617506076158e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2346989621910721e+05, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6261150905514889e-04, + "cpu_time": 5.5051252014770146e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6255025927194149e-04, + "own_numa": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 52641, + "real_time": 1.3303220245301385e+04, + "cpu_time": 2.1281443361638259e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4631630083380342e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 52641, + "real_time": 1.3298491159102583e+04, + "cpu_time": 2.1285288216409230e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4640389355427651e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 52641, + "real_time": 1.3299138695366622e+04, + "cpu_time": 2.1259317167227116e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4639189612644820e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 52641, + "real_time": 1.3301007920854543e+04, + "cpu_time": 2.1301811477745461e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4635727002780981e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 52641, + "real_time": 1.3299730704492045e+04, + "cpu_time": 2.1309309093672269e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4638092851708984e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3300317745023436e+04, + "cpu_time": 2.1287433863338469e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4637005781188555e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3299730704492049e+04, + "cpu_time": 2.1285288216409233e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 2.4638092851708984e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8689596991158961e+00, + "cpu_time": 1.9469720142743476e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4616765602811595e+05, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4051992854194800e-04, + "cpu_time": 9.1461095159405361e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4050719438172567e-04, + "own_numa": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51034, + "real_time": 1.3720347545104036e+04, + "cpu_time": 2.3503861837206565e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7765553885977068e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51034, + "real_time": 1.3720279803984004e+04, + "cpu_time": 2.3506104361797970e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7765789718785543e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51034, + "real_time": 1.3734347556406927e+04, + "cpu_time": 2.3491187482854628e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7716864402072134e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51034, + "real_time": 1.3723028327251232e+04, + "cpu_time": 2.3482363071677748e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7756222924832420e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51034, + "real_time": 1.3715694856741407e+04, + "cpu_time": 2.3495123211976304e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7781757092524090e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3722739617897523e+04, + "cpu_time": 2.3495727993102642e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7757237604838257e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3720347545104034e+04, + "cpu_time": 2.3495123211976308e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 4.7765553885977068e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0040871259057820e+00, + "cpu_time": 9.6621887475303758e+00, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4364572323601330e+06, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.1040006011415437e-04, + "cpu_time": 4.1123172477851245e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1017549476381290e-04, + "own_numa": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 38978, + "real_time": 1.7968485754897763e+04, + "cpu_time": 3.1748902149930731e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.2945490114142218e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 38978, + "real_time": 1.7958427993333866e+04, + "cpu_time": 3.1733720252449948e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.2986343820658293e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 38978, + "real_time": 1.7973468434851544e+04, + "cpu_time": 3.1713126943403811e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.2925267860845480e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 38978, + "real_time": 1.7970139272219414e+04, + "cpu_time": 3.1765658627944067e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.2938778055342169e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 38978, + "real_time": 1.7960823707052594e+04, + "cpu_time": 3.1759489814767414e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.2976608499604940e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7966269032471035e+04, + "cpu_time": 3.1744179557699194e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.2954497670118628e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7968485754897767e+04, + "cpu_time": 3.1748902149930735e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 7.2945490114142218e+09, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3807811375348384e+00, + "cpu_time": 2.1164352273604386e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5911822512513474e+06, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.5515337803317088e-04, + "cpu_time": 6.6671599545155714e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5517786209261608e-04, + "own_numa": NaN, + "wr_gpu": NaN + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 30312, + "real_time": 2.3097334131614200e+04, + "cpu_time": 4.5072066640274476e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1349534907632198e+10, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 30312, + "real_time": 2.3098034811769248e+04, + "cpu_time": 4.5092631730007786e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1349190618867218e+10, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 30312, + "real_time": 2.3092193061830068e+04, + "cpu_time": 4.5194068355766532e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1352061681543249e+10, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 30312, + "real_time": 2.3094934982585437e+04, + "cpu_time": 4.5199293910002714e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1350713920505415e+10, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 30312, + "real_time": 2.3095478893421987e+04, + "cpu_time": 4.5219164258379526e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1350446605143288e+10, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3095595176244191e+04, + "cpu_time": 4.5155444978886211e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1350389546738274e+10, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3095478893421987e+04, + "cpu_time": 4.5194068355766540e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1350446605143288e+10, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2912631354625681e+00, + "cpu_time": 6.7771845493547289e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1260933400744363e+06, + "own_numa": 0.0000000000000000e+00, + "wr_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_implicit_mapped_GPUWrHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.9207797763070001e-05, + "cpu_time": 1.5008565528528413e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9211867173143700e-05, + "own_numa": NaN, + "wr_gpu": NaN + } \ No newline at end of file diff --git a/results/implicit_mapped_HostWrGPU.json b/results/implicit_mapped_HostWrGPU.json new file mode 100644 index 0000000..3a89567 --- /dev/null +++ b/results/implicit_mapped_HostWrGPU.json @@ -0,0 +1,38 @@ +{ + "context": { + "date": "2023-10-09T17:44:34-04:00", + "host_name": "frontier10439", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.39,3.82,18.5], + "library_build_type": "release" + }, + "benchmarks": [ diff --git a/results/lscpu.r1.1465528.txt b/results/lscpu.r1.1465528.txt new file mode 100644 index 0000000..9380a1b --- /dev/null +++ b/results/lscpu.r1.1465528.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.40 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r10.1465527.txt b/results/lscpu.r10.1465527.txt new file mode 100644 index 0000000..d1b5d89 --- /dev/null +++ b/results/lscpu.r10.1465527.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.17 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r11.1465787.txt b/results/lscpu.r11.1465787.txt new file mode 100644 index 0000000..7e89dad --- /dev/null +++ b/results/lscpu.r11.1465787.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.34 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r12.1467862.txt b/results/lscpu.r12.1467862.txt new file mode 100644 index 0000000..c0583dc --- /dev/null +++ b/results/lscpu.r12.1467862.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.35 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r2.1465529.txt b/results/lscpu.r2.1465529.txt new file mode 100644 index 0000000..13c65f3 --- /dev/null +++ b/results/lscpu.r2.1465529.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.46 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r3.1465530.txt b/results/lscpu.r3.1465530.txt new file mode 100644 index 0000000..74a50f9 --- /dev/null +++ b/results/lscpu.r3.1465530.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.55 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r4.1465531.txt b/results/lscpu.r4.1465531.txt new file mode 100644 index 0000000..2ebbba1 --- /dev/null +++ b/results/lscpu.r4.1465531.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.45 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r5.1465786.txt b/results/lscpu.r5.1465786.txt new file mode 100644 index 0000000..f8e17d8 --- /dev/null +++ b/results/lscpu.r5.1465786.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.69 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r6.1465533.txt b/results/lscpu.r6.1465533.txt new file mode 100644 index 0000000..f81011b --- /dev/null +++ b/results/lscpu.r6.1465533.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.47 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r7.1465534.txt b/results/lscpu.r7.1465534.txt new file mode 100644 index 0000000..8b7c570 --- /dev/null +++ b/results/lscpu.r7.1465534.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.39 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r8.1465535.txt b/results/lscpu.r8.1465535.txt new file mode 100644 index 0000000..2ebbba1 --- /dev/null +++ b/results/lscpu.r8.1465535.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.45 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/lscpu.r9.1465536.txt b/results/lscpu.r9.1465536.txt new file mode 100644 index 0000000..74a50f9 --- /dev/null +++ b/results/lscpu.r9.1465536.txt @@ -0,0 +1,40 @@ +Architecture: x86_64 +CPU op-mode(s): 32-bit, 64-bit +Address sizes: 48 bits physical, 48 bits virtual +Byte Order: Little Endian +CPU(s): 128 +On-line CPU(s) list: 0-127 +Vendor ID: AuthenticAMD +Model name: AMD EPYC 7A53 64-Core Processor +CPU family: 25 +Model: 48 +Thread(s) per core: 2 +Core(s) per socket: 64 +Socket(s): 1 +Stepping: 1 +Frequency boost: enabled +CPU max MHz: 2000.0000 +CPU min MHz: 1500.0000 +BogoMIPS: 3992.55 +Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm +Virtualization: AMD-V +L1d cache: 2 MiB (64 instances) +L1i cache: 2 MiB (64 instances) +L2 cache: 32 MiB (64 instances) +L3 cache: 256 MiB (8 instances) +NUMA node(s): 4 +NUMA node0 CPU(s): 0-15,64-79 +NUMA node1 CPU(s): 16-31,80-95 +NUMA node2 CPU(s): 32-47,96-111 +NUMA node3 CPU(s): 48-63,112-127 +Vulnerability Itlb multihit: Not affected +Vulnerability L1tf: Not affected +Vulnerability Mds: Not affected +Vulnerability Meltdown: Not affected +Vulnerability Mmio stale data: Not affected +Vulnerability Retbleed: Not affected +Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp +Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization +Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected +Vulnerability Srbds: Not affected +Vulnerability Tsx async abort: Not affected diff --git a/results/modules.r1.1465528.txt b/results/modules.r1.1465528.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r1.1465528.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r10.1465527.txt b/results/modules.r10.1465527.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r10.1465527.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r11.1465787.txt b/results/modules.r11.1465787.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r11.1465787.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r12.1467862.txt b/results/modules.r12.1467862.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r12.1467862.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r2.1465529.txt b/results/modules.r2.1465529.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r2.1465529.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r3.1465530.txt b/results/modules.r3.1465530.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r3.1465530.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r4.1465531.txt b/results/modules.r4.1465531.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r4.1465531.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r5.1465786.txt b/results/modules.r5.1465786.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r5.1465786.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r6.1465533.txt b/results/modules.r6.1465533.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r6.1465533.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r7.1465534.txt b/results/modules.r7.1465534.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r7.1465534.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r8.1465535.txt b/results/modules.r8.1465535.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r8.1465535.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/modules.r9.1465536.txt b/results/modules.r9.1465536.txt new file mode 100644 index 0000000..e087a59 --- /dev/null +++ b/results/modules.r9.1465536.txt @@ -0,0 +1,13 @@ + +Currently Loaded Modules: + 1) craype-x86-trento 9) craype/2.7.19 + 2) libfabric/1.15.2.0 10) cray-dsmml/0.2.2 + 3) craype-network-ofi 11) cray-mpich/8.1.23 + 4) perftools-base/22.12.0 12) cray-libsci/22.12.1.1 + 5) xpmem/2.6.2-2.5_2.22__gd067c3f.shasta 13) PrgEnv-amd/8.3.3 + 6) cray-pmi/6.1.8 14) darshan-runtime/3.4.0 + 7) cray-python/3.9.13.1 15) hsi/default + 8) amd/5.3.0 16) DefApps/default + + + diff --git a/results/prefetch_managed_GPUToGPU.json b/results/prefetch_managed_GPUToGPU.json new file mode 100644 index 0000000..71a9276 --- /dev/null +++ b/results/prefetch_managed_GPUToGPU.json @@ -0,0 +1,27928 @@ +{ + "context": { + "date": "2023-10-09T17:44:34-04:00", + "host_name": "frontier07936", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1796, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.1,3.44,18.39], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 78707, + "real_time": 8.8946550371526737e+03, + "cpu_time": 2.1083203260192869e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1512531916333874e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 78707, + "real_time": 8.9399018746996080e+03, + "cpu_time": 2.1139228848768224e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1454264424288300e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 78707, + "real_time": 8.8822038082539020e+03, + "cpu_time": 2.1130249012159020e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1528670385253206e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 78707, + "real_time": 8.8197218299117012e+03, + "cpu_time": 2.0986276087260343e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1610343497763713e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 78707, + "real_time": 8.8473215126098112e+03, + "cpu_time": 2.1024858868969724e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1574124423313031e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8767608125255392e+03, + "cpu_time": 2.1072763215470040e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1535986929390424e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8822038082539020e+03, + "cpu_time": 2.1083203260192873e+04, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.1528670385253206e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5980270570961792e+01, + "cpu_time": 6.6344068166224488e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9719354062983638e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.1798478681639596e-03, + "cpu_time": 3.1483326362021502e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1767875976727797e-03, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 78589, + "real_time": 8.9634397051780197e+03, + "cpu_time": 2.1176430543714781e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.2848371466334593e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 78589, + "real_time": 8.9726139456816327e+03, + "cpu_time": 2.1190018374072701e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.2825009661601099e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 78589, + "real_time": 8.9018022576698768e+03, + "cpu_time": 2.1125435328099349e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.3006577103366053e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 78589, + "real_time": 8.7360672893992414e+03, + "cpu_time": 2.0905029037142616e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.3443042872221696e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 78589, + "real_time": 8.7523678634582247e+03, + "cpu_time": 2.0947902174604566e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.3399382109503755e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8652582122774002e+03, + "cpu_time": 2.1068963091526803e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.3104476642605439e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9018022576698750e+03, + "cpu_time": 2.1125435328099345e+04, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.3006577103366053e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1394609658430871e+02, + "cpu_time": 1.3315673847766939e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9785579598641354e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2853105217680632e-02, + "cpu_time": 6.3200423247796343e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2891691969215928e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 76563, + "real_time": 9.1758126410856839e+03, + "cpu_time": 2.1678555020048840e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.4639098031053090e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 76563, + "real_time": 9.1749560630861542e+03, + "cpu_time": 2.1661326293379272e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.4643265557200283e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 76563, + "real_time": 8.8873636969904182e+03, + "cpu_time": 2.1143602823818313e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.6087907951680350e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 76563, + "real_time": 8.8058834564940189e+03, + "cpu_time": 2.0943220994475148e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.6514356228270864e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 76563, + "real_time": 8.7819553635613138e+03, + "cpu_time": 2.0899924937633055e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.6641093360544759e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9651942442435175e+03, + "cpu_time": 2.1265326013870923e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.5705144225749874e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.8873636969904201e+03, + "cpu_time": 2.1143602823818313e+04, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 4.6087907951680350e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9581505196566229e+02, + "cpu_time": 3.8067721167339243e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9264466373702940e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1841696524466676e-02, + "cpu_time": 1.7901310867516666e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1718445058046272e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 74859, + "real_time": 9.3583584341838359e+03, + "cpu_time": 2.1749885504748938e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 8.7536719795606375e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 74859, + "real_time": 9.3349202606052932e+03, + "cpu_time": 2.1734048491163372e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 8.7756507514814234e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 74859, + "real_time": 9.0723608122404366e+03, + "cpu_time": 2.1466648004915914e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 9.0296232364869642e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 74859, + "real_time": 8.8957468149671076e+03, + "cpu_time": 2.1171993400927040e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 9.2088951837263930e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 74859, + "real_time": 8.8867957603345749e+03, + "cpu_time": 2.1153040168850766e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 9.2181706668271434e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1096364164662500e+03, + "cpu_time": 2.1455123114121212e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 8.9972023636165130e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0723608122404366e+03, + "cpu_time": 2.1466648004915918e+04, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 9.0296232364869642e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2880758447473301e+02, + "cpu_time": 2.8992826539989221e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2532427682427831e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5117092934812263e-02, + "cpu_time": 1.3513241749196436e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5043815590439496e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 74783, + "real_time": 9.3233176343006526e+03, + "cpu_time": 2.1725976438495378e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.7573143641189454e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 74783, + "real_time": 9.2636954546162724e+03, + "cpu_time": 2.1670255566104588e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.7686246358450339e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 74783, + "real_time": 8.9396207907288917e+03, + "cpu_time": 2.1253246326036617e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.8327399319881141e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 74783, + "real_time": 8.8855296691966942e+03, + "cpu_time": 2.1106793469103930e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.8438968311363721e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 74783, + "real_time": 8.8923117554261426e+03, + "cpu_time": 2.1164502440394230e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.8424905076008365e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0608950608537307e+03, + "cpu_time": 2.1384154848026952e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.8090132541378603e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9396207907288936e+03, + "cpu_time": 2.1253246326036620e+04, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 1.8327399319881141e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1440289574711588e+02, + "cpu_time": 2.9197995497163049e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2439843903577216e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3662441106222731e-02, + "cpu_time": 1.3654032953215852e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3460217224225480e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 74615, + "real_time": 9.4267228447175094e+03, + "cpu_time": 2.1864759337934756e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.4760754654373164e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 74615, + "real_time": 9.3264794118711234e+03, + "cpu_time": 2.1722734704818133e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.5134372310189791e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 74615, + "real_time": 9.0245445681393903e+03, + "cpu_time": 2.1330328459425029e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.6309865558961773e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 74615, + "real_time": 8.9498231561603916e+03, + "cpu_time": 2.1232200254640429e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.6613013942565942e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 74615, + "real_time": 8.9232025874020965e+03, + "cpu_time": 2.1201802305166533e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.6722241458758678e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1301545136581044e+03, + "cpu_time": 2.1470365012396978e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.5908049584969864e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0245445681393903e+03, + "cpu_time": 2.1330328459425029e+04, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 3.6309865558961773e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3075863987950260e+02, + "cpu_time": 3.0318969366947397e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9947226375604928e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5274341144424558e-02, + "cpu_time": 1.4121310629531094e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5049321089623981e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 74526, + "real_time": 9.3427494888606434e+03, + "cpu_time": 2.1732281713764369e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.0146374017775545e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 74526, + "real_time": 9.3523767887819886e+03, + "cpu_time": 2.1789629001959071e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.0074165615963297e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 74526, + "real_time": 9.0842728629241174e+03, + "cpu_time": 2.1513722955746936e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.2142262775344200e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 74526, + "real_time": 9.0215619077872107e+03, + "cpu_time": 2.1311267450285886e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.2643740263457918e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 74526, + "real_time": 8.9296963520675727e+03, + "cpu_time": 2.1205923811824097e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.3391073353603868e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1461314800843065e+03, + "cpu_time": 2.1510564986716072e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.1679523205228958e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0842728629241155e+03, + "cpu_time": 2.1513722955746933e+04, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.2142262775344200e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9195311942884425e+02, + "cpu_time": 2.5474043548871538e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5000804791695037e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0987356222335316e-02, + "cpu_time": 1.1842572970353465e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0927601246377631e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 74998, + "real_time": 9.3520810332323872e+03, + "cpu_time": 2.1764504680124803e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.4015276336276268e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 74998, + "real_time": 9.3219381380641007e+03, + "cpu_time": 2.1734253566761814e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.4060595346025318e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 74998, + "real_time": 8.9867783086506970e+03, + "cpu_time": 2.1310168044481092e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.4584982014503433e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 74998, + "real_time": 8.8710032554500849e+03, + "cpu_time": 2.1134989839729020e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.4775329940215410e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 74998, + "real_time": 8.8922885050150417e+03, + "cpu_time": 2.1167287260993515e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.4739962600862364e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0848178480824645e+03, + "cpu_time": 2.1422240678418049e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.4435229247576561e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9867783086506970e+03, + "cpu_time": 2.1310168044481092e+04, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.4584982014503433e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3454830063015939e+02, + "cpu_time": 3.0601113297496829e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7002290362943447e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.5817611817023457e-02, + "cpu_time": 1.4284739751022440e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5633323675240922e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 76516, + "real_time": 9.1420993720414808e+03, + "cpu_time": 2.1638188548800255e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.8674376566250542e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 76516, + "real_time": 9.1997741953995483e+03, + "cpu_time": 2.1715624536044728e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.8494612414627319e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 76516, + "real_time": 9.0976966306127324e+03, + "cpu_time": 2.1653628809660804e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.8814326377724525e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 76516, + "real_time": 8.9180091048972899e+03, + "cpu_time": 2.1232862734591479e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.9394901588073578e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 76516, + "real_time": 8.8703865766987801e+03, + "cpu_time": 2.1209333328976896e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.9552714273875538e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0455931759299674e+03, + "cpu_time": 2.1489927591614840e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.8986186244110306e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0976966306127324e+03, + "cpu_time": 2.1638188548800259e+04, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.8814326377724525e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4385326746188511e+02, + "cpu_time": 2.4725189174292797e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6271161634806275e+08, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5903132571192129e-02, + "cpu_time": 1.1505478121731935e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5963176819857802e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 74850, + "real_time": 9.4167047704050874e+03, + "cpu_time": 2.1845321389445551e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 5.5676376480203300e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 74850, + "real_time": 9.3964081119430866e+03, + "cpu_time": 2.1794509311957241e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 5.5796639923889206e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 74850, + "real_time": 9.0883293987845209e+03, + "cpu_time": 2.1533765691382654e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 5.7688049914885193e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 74850, + "real_time": 8.9647603087111202e+03, + "cpu_time": 2.1238487227788770e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 5.8483214491584984e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 74850, + "real_time": 8.8936077164248745e+03, + "cpu_time": 2.1160328122912517e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 5.8951104739164001e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1519620612537365e+03, + "cpu_time": 2.1514482348697351e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 5.7319077109945343e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0883293987845209e+03, + "cpu_time": 2.1533765691382654e+04, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 5.7688049914885193e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4273604652293261e+02, + "cpu_time": 3.1218581836266185e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5141954382680483e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6522842304011900e-02, + "cpu_time": 1.4510496385778204e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6416954260509589e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 75358, + "real_time": 9.3604804976560736e+03, + "cpu_time": 2.1865423219830678e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.1202159977391870e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 75358, + "real_time": 9.3105925958204516e+03, + "cpu_time": 2.1823221277103956e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.1262183252124127e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 75358, + "real_time": 9.0990085645169493e+03, + "cpu_time": 2.1614953846970355e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.1524068722048369e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 75358, + "real_time": 8.8920673793312963e+03, + "cpu_time": 2.1143430266196105e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.1792263320420940e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 75358, + "real_time": 8.8778081139068800e+03, + "cpu_time": 2.1171901058945255e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.1811203695171446e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1079914302463312e+03, + "cpu_time": 2.1523785933809275e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.1518375793431351e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0990085645169493e+03, + "cpu_time": 2.1614953846970355e+04, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.1524068722048369e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2610075641997531e+02, + "cpu_time": 3.4755666712316952e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8566677591607466e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4824436666587891e-02, + "cpu_time": 1.6147561966653467e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4800959878300156e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 76558, + "real_time": 9.1357539675155858e+03, + "cpu_time": 2.1559371025888824e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.2955434301940906e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 76558, + "real_time": 9.1438940866444227e+03, + "cpu_time": 2.1550364625512717e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.2934998810442279e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 76558, + "real_time": 8.9897869125182842e+03, + "cpu_time": 2.1395398782622320e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3328161394790286e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 76558, + "real_time": 8.9709263659188146e+03, + "cpu_time": 2.1276926944277511e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3377206705956577e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 76558, + "real_time": 8.9708858362032879e+03, + "cpu_time": 2.1277738956085635e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3377312322230701e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0422494337600820e+03, + "cpu_time": 2.1411960066877400e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3194622707072150e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9897869125182842e+03, + "cpu_time": 2.1395398782622320e+04, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.3328161394790286e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9452211390144925e+01, + "cpu_time": 1.3911266980827199e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2867042556015997e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.8926945164957251e-03, + "cpu_time": 6.4969610149548246e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.8587689245075438e-03, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 75798, + "real_time": 9.2478302218325389e+03, + "cpu_time": 2.1676233040449621e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.5354465851870514e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 75798, + "real_time": 9.1974185520243209e+03, + "cpu_time": 2.1629489973350326e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.5603056730269684e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 75798, + "real_time": 9.2196018981317047e+03, + "cpu_time": 2.1637845536821518e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.5493330908896942e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 75798, + "real_time": 8.9916270023185662e+03, + "cpu_time": 2.1360696561914476e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.6646774815263837e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 75798, + "real_time": 8.9811736664209729e+03, + "cpu_time": 2.1294938942979890e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.6701067764470074e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1275302681456233e+03, + "cpu_time": 2.1519840811103168e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.5959739214154211e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1974185520243227e+03, + "cpu_time": 2.1629489973350326e+04, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 4.5603056730269684e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3011878098455850e+02, + "cpu_time": 1.7770363376152292e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5816048219992971e+09, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4255639495238162e-02, + "cpu_time": 8.2576648833683139e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4320370251301081e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 75935, + "real_time": 9.2252524073515433e+03, + "cpu_time": 2.1646001422269201e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 9.0930932072006763e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 75935, + "real_time": 9.1976453694550819e+03, + "cpu_time": 2.1638607401066802e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 9.1203864283114734e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 75935, + "real_time": 9.1497301000625303e+03, + "cpu_time": 2.1534954447883214e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 9.1681480308830872e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 75935, + "real_time": 8.8689514220367964e+03, + "cpu_time": 2.1259638743662344e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 9.4583988577913708e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 75935, + "real_time": 8.9008076274023424e+03, + "cpu_time": 2.1294042207150887e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 9.4245470199519116e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0684773852616599e+03, + "cpu_time": 2.1474648844406493e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 9.2529147088277039e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.1497301000625303e+03, + "cpu_time": 2.1534954447883214e+04, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 9.1681480308830872e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7013857370957450e+02, + "cpu_time": 1.8623169768837948e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7462347558808338e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8761536968277436e-02, + "cpu_time": 8.6721649810300536e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8872266856786717e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 76383, + "real_time": 9.1832435213790322e+03, + "cpu_time": 2.1612341083748903e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8269379398403010e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 76383, + "real_time": 9.1827407255110447e+03, + "cpu_time": 2.1636244229737047e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8270379728124475e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 76383, + "real_time": 8.9516124896869806e+03, + "cpu_time": 2.1378866789730564e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8742116037003142e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 76383, + "real_time": 8.8851413449603642e+03, + "cpu_time": 2.1190325046149122e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8882328765108511e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 76383, + "real_time": 8.8979023380848339e+03, + "cpu_time": 2.1220543655002071e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8855248532218770e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0201280839244500e+03, + "cpu_time": 2.1407664160873541e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8603890492171582e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9516124896869806e+03, + "cpu_time": 2.1378866789730564e+04, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 1.8742116037003142e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5075182953759051e+02, + "cpu_time": 2.1048821447474683e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0941159438760578e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6712825819653089e-02, + "cpu_time": 9.8323765214634162e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6631553196779166e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 77408, + "real_time": 9.0485900540925541e+03, + "cpu_time": 2.1471196517155884e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7082497714463027e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 77408, + "real_time": 9.0664676877888887e+03, + "cpu_time": 2.1507375620090934e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7009376921061064e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 77408, + "real_time": 8.9627881161221849e+03, + "cpu_time": 2.1461209306531713e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7437493294795825e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 77408, + "real_time": 8.9163154845582430e+03, + "cpu_time": 2.1371139520462992e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7632620848949746e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 77408, + "real_time": 8.8733104606546858e+03, + "cpu_time": 2.1336832226643255e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7815009571438237e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9734943606433117e+03, + "cpu_time": 2.1429550638176956e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7395399670141582e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9627881161221849e+03, + "cpu_time": 2.1461209306531713e+04, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.7437493294795825e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.3222989753685056e+01, + "cpu_time": 7.2114389753522843e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4678502954016357e+10, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.2743123702948189e-03, + "cpu_time": 3.3651844115223932e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.2734676617737734e-03, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 77276, + "real_time": 9.1036768586938288e+03, + "cpu_time": 2.1643304027123522e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.3716219327262676e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 77276, + "real_time": 9.1521969718240925e+03, + "cpu_time": 2.1712518776851830e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.3325414877543633e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 77276, + "real_time": 9.0714718640600877e+03, + "cpu_time": 2.1586477017444118e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.3977922222165508e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 77276, + "real_time": 8.8842184367849259e+03, + "cpu_time": 2.1442871952481975e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.5537161178002012e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 77276, + "real_time": 8.8302668836560151e+03, + "cpu_time": 2.1358033270355627e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.5998681448928945e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0083662030037904e+03, + "cpu_time": 2.1548641008851413e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.4511079810780566e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0714718640600877e+03, + "cpu_time": 2.1586477017444115e+04, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.3977922222165508e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4220211937067518e+02, + "cpu_time": 1.4557384438780949e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1819119195640260e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5785561573113966e-02, + "cpu_time": 6.7555928157145935e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5862230457073877e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 76837, + "real_time": 9.1145764790523281e+03, + "cpu_time": 2.1733090073792402e+04, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.4725613231560164e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 76837, + "real_time": 9.1211666827156350e+03, + "cpu_time": 2.1792262816091432e+04, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.4714973716502625e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 76837, + "real_time": 8.9649375985596635e+03, + "cpu_time": 2.1551706781889021e+04, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.4971406830713900e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 76837, + "real_time": 8.8778236001933783e+03, + "cpu_time": 2.1450491247706388e+04, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.5118314357707721e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 76837, + "real_time": 8.8262126208425198e+03, + "cpu_time": 2.1364456108385468e+04, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.5206718188847352e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9809433962727053e+03, + "cpu_time": 2.1578401405572942e+04, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.4947405265066352e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9649375985596635e+03, + "cpu_time": 2.1551706781889025e+04, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 1.4971406830713900e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3449150683509407e+02, + "cpu_time": 1.8201063745846497e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2374389178015207e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4975209273773077e-02, + "cpu_time": 8.4348527046798766e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4968744595629914e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 76033, + "real_time": 9.2351797342299687e+03, + "cpu_time": 2.2189862138808116e+04, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.9066619570494172e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 76033, + "real_time": 9.1274142838408206e+03, + "cpu_time": 2.2077648008101845e+04, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.9409803001408434e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 76033, + "real_time": 8.9842026476551018e+03, + "cpu_time": 2.1916466310680713e+04, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.9878606541679273e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 76033, + "real_time": 8.8929258204204689e+03, + "cpu_time": 2.1733486209935189e+04, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.0185280010275406e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 76033, + "real_time": 8.9517912889184172e+03, + "cpu_time": 2.1860166072626085e+04, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.9986786704053422e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0383027550129573e+03, + "cpu_time": 2.1955525748030392e+04, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.9705419165582141e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9842026476551036e+03, + "cpu_time": 2.1916466310680713e+04, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 2.9878606541679273e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3986325286606814e+02, + "cpu_time": 1.8010278611448967e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5685934061240118e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5474504080812641e-02, + "cpu_time": 8.2030732573391690e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5379663153911535e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 76789, + "real_time": 9.1097676811269575e+03, + "cpu_time": 2.2585457044628962e+04, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.8933545924805016e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 76789, + "real_time": 9.1654155435420962e+03, + "cpu_time": 2.2681542994439333e+04, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.8575730631032484e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 76789, + "real_time": 9.2970014330258928e+03, + "cpu_time": 2.2933462930888421e+04, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.7746674115039336e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 76789, + "real_time": 9.2245722952099895e+03, + "cpu_time": 2.2725995324851381e+04, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.8200087203910695e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 76789, + "real_time": 9.2202419306079937e+03, + "cpu_time": 2.2718884358436502e+04, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.8227421367087500e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.2033997767025885e+03, + "cpu_time": 2.2729068530648925e+04, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.8336691848375000e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.2202419306079955e+03, + "cpu_time": 2.2718884358436502e+04, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.8227421367087500e+13, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0183643117802035e+01, + "cpu_time": 1.2725286668229396e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4499891033343146e+11, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.6258388009466174e-03, + "cpu_time": 5.5986837520728278e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6281135634163862e-03, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 74438, + "real_time": 9.3404369071041401e+03, + "cpu_time": 2.4037771998173288e+04, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1495627396009009e+14, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 74438, + "real_time": 9.6937157363412043e+03, + "cpu_time": 2.4539952604852406e+04, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1076679502521425e+14, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 74438, + "real_time": 9.6122363763372596e+03, + "cpu_time": 2.4628416803245920e+04, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1170572403351042e+14, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 74438, + "real_time": 9.3672727632406368e+03, + "cpu_time": 2.4133849163061834e+04, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1462694117476897e+14, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 74438, + "real_time": 8.9507168330968270e+03, + "cpu_time": 2.3556344031274453e+04, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1996154542948488e+14, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.3928757232240132e+03, + "cpu_time": 2.4179266920121579e+04, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1440345592461373e+14, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.3672727632406368e+03, + "cpu_time": 2.4133849163061830e+04, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 1.1462694117476897e+14, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9048822469297579e+02, + "cpu_time": 4.3068152146433670e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5969074298288706e+12, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.0926441832370859e-02, + "cpu_time": 1.7812017332334042e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1440548720827598e-02, + "dst_gpu": NaN, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2687, + "real_time": 2.6025346192716027e+05, + "cpu_time": 5.2394807927055238e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9346258544164808e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2687, + "real_time": 2.6055858456852016e+05, + "cpu_time": 5.2384401898027392e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9300182785983565e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2687, + "real_time": 2.6031111108047655e+05, + "cpu_time": 5.2332412095273857e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9337544822795712e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2687, + "real_time": 2.5984726328566353e+05, + "cpu_time": 5.2356186416077410e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9407765433121524e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2687, + "real_time": 2.6108538687307487e+05, + "cpu_time": 5.2299073018236260e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9220885253827381e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6041116154697910e+05, + "cpu_time": 5.2353376270934037e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9322527367978599e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time_median", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6031111108047655e+05, + "cpu_time": 5.2356186416077410e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9337544822795712e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5529491890567954e+02, + "cpu_time": 3.8972453398814071e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8710661717752409e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7483694485328069e-03, + "cpu_time": 7.4441146254116775e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7473612790643096e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2675, + "real_time": 2.6058344817807394e+05, + "cpu_time": 5.2445306915887212e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8592865906067295e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2675, + "real_time": 2.5992927241583014e+05, + "cpu_time": 5.2253890205607412e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8790664128188174e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2675, + "real_time": 2.6011780453802340e+05, + "cpu_time": 5.2353837084112212e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8733557037254954e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2675, + "real_time": 2.5967907962371883e+05, + "cpu_time": 5.2325564747663826e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8866576505415868e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2675, + "real_time": 2.6103869585372577e+05, + "cpu_time": 5.2477074691588548e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8455801095007239e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6026966012187442e+05, + "cpu_time": 5.2371134728971851e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8687892934386702e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time_median", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6011780453802337e+05, + "cpu_time": 5.2353837084112206e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8733557037254954e+06, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4258366512257226e+02, + "cpu_time": 9.0618567306414991e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6391438705969726e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0846980968450214e-03, + "cpu_time": 1.7303151397306761e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0830953904989679e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2684, + "real_time": 2.6043657332380483e+05, + "cpu_time": 5.2437657861400151e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5727437770068415e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2684, + "real_time": 2.6160010054947642e+05, + "cpu_time": 5.2514349962741858e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5657486336574718e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2684, + "real_time": 2.6013362215130607e+05, + "cpu_time": 5.2433610543963499e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5745753917260151e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2684, + "real_time": 2.6070877291626835e+05, + "cpu_time": 5.2412137146049843e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5711017140629591e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2684, + "real_time": 2.6054530797778504e+05, + "cpu_time": 5.2340492138599302e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5720874161162168e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6068487538372818e+05, + "cpu_time": 5.2427649530550942e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5712513865139008e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time_median", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6054530797778504e+05, + "cpu_time": 5.2433610543963511e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5720874161162168e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.5299667629051100e+02, + "cpu_time": 6.2228832455899578e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3270811671747280e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1213224414208909e-03, + "cpu_time": 1.1869468307870116e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1174722235608946e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2707, + "real_time": 2.6167014500514665e+05, + "cpu_time": 5.2685423642409418e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1306590210506726e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2707, + "real_time": 2.6276437094346335e+05, + "cpu_time": 5.2730169782046322e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1176220621488292e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2707, + "real_time": 2.6107086589269884e+05, + "cpu_time": 5.2794092944220023e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1378453401870374e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2707, + "real_time": 2.6287514162736683e+05, + "cpu_time": 5.2788007129664149e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1163083543336321e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2707, + "real_time": 2.6182772012084781e+05, + "cpu_time": 5.2728563502032764e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1287749044367585e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6204164871790475e+05, + "cpu_time": 5.2745251400074537e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1262419364313859e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time_median", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6182772012084784e+05, + "cpu_time": 5.2730169782046322e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1287749044367585e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6538186422445574e+02, + "cpu_time": 4.5549013647087958e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.1315911554168648e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9208405151213614e-03, + "cpu_time": 8.6356614933156964e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9209483274479395e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2652, + "real_time": 2.6414604477589979e+05, + "cpu_time": 5.3209599849171273e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2026293120913863e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2652, + "real_time": 2.6560812602701091e+05, + "cpu_time": 5.3276899585217331e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1684859740826741e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2652, + "real_time": 2.6390972852055123e+05, + "cpu_time": 5.3303321003018273e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2081834162942357e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2652, + "real_time": 2.6519733353719523e+05, + "cpu_time": 5.3293767835596064e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1780410011935741e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2652, + "real_time": 2.6479727270076022e+05, + "cpu_time": 5.3300968099547015e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1873749049202211e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6473170111228351e+05, + "cpu_time": 5.3276911274509993e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1889429217164189e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time_median", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6479727270076016e+05, + "cpu_time": 5.3293767835596076e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1873749049202211e+07, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.0849596071908013e+02, + "cpu_time": 3.9022994375804001e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6562843711905272e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6762792583672410e-03, + "cpu_time": 7.3245601973315427e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6761991379477437e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2604, + "real_time": 2.6853901914216450e+05, + "cpu_time": 5.4155939016897883e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2202323559785040e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2604, + "real_time": 2.6930137085662183e+05, + "cpu_time": 5.4216672542242566e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2167780615363427e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2604, + "real_time": 2.6966489317859831e+05, + "cpu_time": 5.4266058333332685e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2151377813313594e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2604, + "real_time": 2.6880336844066542e+05, + "cpu_time": 5.4198261712749570e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2190323428641511e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2604, + "real_time": 2.7006577207757294e+05, + "cpu_time": 5.4187415322579257e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2133340611037450e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6927488473912462e+05, + "cpu_time": 5.4204869385560392e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2169029205628204e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time_median", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6930137085662183e+05, + "cpu_time": 5.4198261712749570e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2167780615363427e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2091397871960567e+02, + "cpu_time": 4.0699780308621621e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8056925903419999e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3058740859592287e-03, + "cpu_time": 7.5085099862750725e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3056009998269709e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2547, + "real_time": 2.7769817910588614e+05, + "cpu_time": 5.6112344601491769e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3599722623680279e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2547, + "real_time": 2.7770236153314624e+05, + "cpu_time": 5.6120554652532761e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3599367192337576e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2547, + "real_time": 2.7689928553020710e+05, + "cpu_time": 5.6046011307421909e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3667811159032637e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2547, + "real_time": 2.7740566954384168e+05, + "cpu_time": 5.5843822889673710e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3624607279211563e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2547, + "real_time": 2.7901180573515332e+05, + "cpu_time": 5.6027841107185057e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3488611826772955e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7774346028964687e+05, + "cpu_time": 5.6030114911661041e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3596024016207004e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time_median", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7769817910588614e+05, + "cpu_time": 5.6046011307421909e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3599722623680279e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8089264225587590e+02, + "cpu_time": 1.1169562620523523e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6205422288510483e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8115608606644280e-03, + "cpu_time": 1.9934927205010784e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8057872056341812e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2382, + "real_time": 2.9070480403071624e+05, + "cpu_time": 5.9057095256088662e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.5087662186054128e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2382, + "real_time": 2.9072388096845074e+05, + "cpu_time": 5.8966810075565265e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.5084703590010172e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2382, + "real_time": 2.9319308734441054e+05, + "cpu_time": 5.8978652728798706e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4705010335400999e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2382, + "real_time": 2.9062950969665870e+05, + "cpu_time": 5.9029026784215390e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.5099343193609250e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2382, + "real_time": 2.9072609475339152e+05, + "cpu_time": 5.8921176364399027e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.5084360284611487e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9119547535872558e+05, + "cpu_time": 5.8990552241813415e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.5012215917937207e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time_median", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9072388096845069e+05, + "cpu_time": 5.8978652728798694e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.5084703590010172e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1173907720302100e+03, + "cpu_time": 5.3447554660776109e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7184138305891280e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.8372532081883803e-03, + "cpu_time": 9.0603584183590092e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8176610405539846e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2154, + "real_time": 3.2612902371240075e+05, + "cpu_time": 6.5556573259053729e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0380457101289320e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2154, + "real_time": 3.2635683978228079e+05, + "cpu_time": 6.5653353481892985e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0324346863660502e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2154, + "real_time": 3.2646031139845232e+05, + "cpu_time": 6.5635203946146567e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0298888056884575e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2154, + "real_time": 3.2565362470768421e+05, + "cpu_time": 6.5355838904365292e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0497798922185433e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2154, + "real_time": 3.2635312286912947e+05, + "cpu_time": 6.5599482358401711e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0325261696705782e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2619058449398953e+05, + "cpu_time": 6.5560090389972052e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0365350528145123e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time_median", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2635312286912941e+05, + "cpu_time": 6.5599482358401699e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0325261696705782e+08, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2362311765821096e+02, + "cpu_time": 1.2002376576135775e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.9806736297132215e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.9212893640151698e-04, + "cpu_time": 1.8307443605922230e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9304906620400709e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1798, + "real_time": 3.8956866861650982e+05, + "cpu_time": 7.8282837875418086e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3458166486076102e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1798, + "real_time": 3.9017342597992573e+05, + "cpu_time": 7.8378634760842822e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3437306722856476e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1798, + "real_time": 3.8972056634751073e+05, + "cpu_time": 7.8313683259177080e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3452921022712886e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1798, + "real_time": 3.8945778645406704e+05, + "cpu_time": 7.8238894771970238e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3461998148080034e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1798, + "real_time": 3.8976211905817670e+05, + "cpu_time": 7.8279281535037479e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3451486800895181e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8973651329123799e+05, + "cpu_time": 7.8298666440489143e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3452375836124139e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time_median", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8972056634751079e+05, + "cpu_time": 7.8282837875418074e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3452921022712886e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7279256880621722e+02, + "cpu_time": 5.2008190436186339e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.4112472712175618e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.9994101015207616e-04, + "cpu_time": 6.6422830426767399e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9959740835854494e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1311, + "real_time": 5.2684517931598041e+05, + "cpu_time": 1.0620939344012381e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9902924828151586e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1311, + "real_time": 5.2931060510865843e+05, + "cpu_time": 1.0614673615560716e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9810220877489226e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1311, + "real_time": 5.2730956087833724e+05, + "cpu_time": 1.0628657254004714e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9885397075929961e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1311, + "real_time": 5.2881864319370047e+05, + "cpu_time": 1.0599177337909965e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9828650398316574e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1311, + "real_time": 5.2779065834004560e+05, + "cpu_time": 1.0636093188405682e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9867270923245902e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2801492936734448e+05, + "cpu_time": 1.0619908147978692e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9858892820626652e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time_median", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2779065834004560e+05, + "cpu_time": 1.0620939344012379e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9867270923245902e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0296709331692059e+03, + "cpu_time": 1.4113225808852899e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8713847635180876e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9500792040159436e-03, + "cpu_time": 1.3289404778457615e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9494464260852611e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 905, + "real_time": 7.7036603132664168e+05, + "cpu_time": 1.5439227215469629e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7222799483883133e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 905, + "real_time": 7.7076063765398122e+05, + "cpu_time": 1.5455369370165274e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7208862227101398e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 905, + "real_time": 7.7046909702694800e+05, + "cpu_time": 1.5444235856353864e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7219157888257909e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 905, + "real_time": 7.7041499554668425e+05, + "cpu_time": 1.5450817116022503e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7221069321370974e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 905, + "real_time": 7.7079122143833595e+05, + "cpu_time": 1.5453266331491808e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7207782622207437e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7056039659851824e+05, + "cpu_time": 1.5448583177900615e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7215934308564167e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time_median", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7046909702694800e+05, + "cpu_time": 1.5450817116022501e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7219157888257909e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0039431329828366e+02, + "cpu_time": 6.6971645131152616e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.0773367555882211e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6006308419545502e-04, + "cpu_time": 4.3351318603091296e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6004386530875635e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 526, + "real_time": 1.3302074109660923e+06, + "cpu_time": 2.6830618992395592e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1531203069706211e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 526, + "real_time": 1.3292200578741389e+06, + "cpu_time": 2.6823879904943425e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1554624647389650e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 526, + "real_time": 1.3294484990404586e+06, + "cpu_time": 2.6809646577946614e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1549202568036871e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 526, + "real_time": 1.3298102229135998e+06, + "cpu_time": 2.6817297072244002e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1540620817384949e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 526, + "real_time": 1.3301517385475961e+06, + "cpu_time": 2.6821082870721379e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1532522782549577e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3297675858683770e+06, + "cpu_time": 2.6820505083650202e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1541634777013454e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time_median", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3298102229135996e+06, + "cpu_time": 2.6821082870721379e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1540620817384949e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3140486274096656e+02, + "cpu_time": 7.7815951102942790e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0233343373893012e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.2442125024370074e-04, + "cpu_time": 2.9013603905013501e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2443921966120631e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 276, + "real_time": 2.5301617966708392e+06, + "cpu_time": 5.0849314239130812e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3154433092135234e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 276, + "real_time": 2.5404604047239907e+06, + "cpu_time": 5.0885988695652662e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3020030481094561e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 276, + "real_time": 2.5303183330992316e+06, + "cpu_time": 5.0831677862318652e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3152382015607138e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 276, + "real_time": 2.5386430544144325e+06, + "cpu_time": 5.0831090615940625e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3043668685178466e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 276, + "real_time": 2.5307839513634858e+06, + "cpu_time": 5.0846794963768572e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3146282579674773e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5340735080543966e+06, + "cpu_time": 5.0848973275362272e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3103359370738039e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time_median", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5307839513634862e+06, + "cpu_time": 5.0846794963768581e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3146282579674773e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0472074236682583e+03, + "cpu_time": 2.2326685036091867e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5880248146223612e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9917367857033431e-03, + "cpu_time": 4.3907838443829036e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9901378409485217e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 140, + "real_time": 4.9580540075632082e+06, + "cpu_time": 9.8990017642859928e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3838308284676576e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 140, + "real_time": 4.9588883395439815e+06, + "cpu_time": 9.9015319571427498e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3832614996011043e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 140, + "real_time": 4.9562449766589059e+06, + "cpu_time": 9.8975819571425710e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3850659277358446e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 140, + "real_time": 4.9559615818517553e+06, + "cpu_time": 9.8990283214285616e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3852594946329927e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 140, + "real_time": 4.9591205082833767e+06, + "cpu_time": 9.9018273571427278e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3831031070885415e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9576538827802455e+06, + "cpu_time": 9.8997942714285217e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3841041715052280e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time_median", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9580540075632073e+06, + "cpu_time": 9.8990283214285616e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3838308284676576e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4734205585506411e+03, + "cpu_time": 1.8208562983311172e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0058191824776459e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9720117486788908e-04, + "cpu_time": 1.8392870077979620e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9721874135755816e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 71, + "real_time": 9.8164474171861801e+06, + "cpu_time": 1.9564316197183520e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4181848660702305e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 71, + "real_time": 9.8363753854179047e+06, + "cpu_time": 1.9643309661971606e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4112598071179061e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 71, + "real_time": 9.8192892496434730e+06, + "cpu_time": 1.9613840169013906e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4171955980641184e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 71, + "real_time": 9.8227235737820752e+06, + "cpu_time": 1.9642507873239975e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4160008421249332e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 71, + "real_time": 9.8035933030113373e+06, + "cpu_time": 1.9588634901408471e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4226666654657326e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8196857858081944e+06, + "cpu_time": 1.9610521760563497e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4170615557685843e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time_median", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8192892496434748e+06, + "cpu_time": 1.9613840169013910e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4171955980641184e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1804481837151721e+04, + "cpu_time": 3.4362619255076708e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1073850850388985e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2021241916123256e-03, + "cpu_time": 1.7522542069318874e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2020225617840950e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1204709780938696e+07, + "cpu_time": 4.2117607151515752e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1648093604339442e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1178949737187587e+07, + "cpu_time": 4.1930105333333910e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1686587310873694e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1222847880739152e+07, + "cpu_time": 4.2080341848485537e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1621045571789074e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1193732705080148e+07, + "cpu_time": 4.1990913515150800e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1664485408893533e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1201363736481376e+07, + "cpu_time": 4.2017415090909213e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1653088373992271e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1200320768085394e+07, + "cpu_time": 4.2027276587879047e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1654660053977604e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time_median", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1201363736481380e+07, + "cpu_time": 4.2017415090909213e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1653088373992271e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6030795030290650e+04, + "cpu_time": 7.3887257865383581e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3934332995427302e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5615813579684799e-04, + "cpu_time": 1.7580786542493493e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5610772488519603e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18, + "real_time": 3.9356644989715680e+07, + "cpu_time": 7.7421940555555895e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4102939423589725e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18, + "real_time": 3.9344075860248670e+07, + "cpu_time": 7.7469421777777836e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4113834183510971e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18, + "real_time": 3.9327062873376742e+07, + "cpu_time": 7.7355495444445908e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4128591914465456e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18, + "real_time": 3.9364742736021675e+07, + "cpu_time": 7.7591955444442332e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4095924086195226e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18, + "real_time": 3.9355862678753003e+07, + "cpu_time": 7.7532199500000045e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4103617317594700e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9349677827623159e+07, + "cpu_time": 7.7474202544444412e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4108981385071220e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time_median", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9355862678753003e+07, + "cpu_time": 7.7469421777777821e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4103617317594700e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4633552101783634e+04, + "cpu_time": 9.2296992242893481e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2687534104939383e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7188492789923163e-04, + "cpu_time": 1.1913254891516402e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7197047785462301e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9, + "real_time": 7.5649515622191966e+07, + "cpu_time": 1.4756960855555689e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5484094483911519e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9, + "real_time": 7.5650668806499898e+07, + "cpu_time": 1.4747835633333278e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5483553580551562e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9, + "real_time": 7.5640661848915949e+07, + "cpu_time": 1.4757690677777824e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5488247913030009e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9, + "real_time": 7.5580374234252512e+07, + "cpu_time": 1.4749481277777931e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5516555550256443e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9, + "real_time": 7.5691077444288462e+07, + "cpu_time": 1.4771848766666710e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5464610237260633e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5642459591229767e+07, + "cpu_time": 1.4756763442222288e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5487412353002038e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time_median", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5649515622191966e+07, + "cpu_time": 1.4756960855555689e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5484094483911519e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9807969831743321e+04, + "cpu_time": 9.5026501830709123e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8680598006702033e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.2626487883742463e-04, + "cpu_time": 6.4395219319446286e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2640068035622091e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5, + "real_time": 1.4327096343040466e+08, + "cpu_time": 2.7409749079999983e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7472415843758221e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5, + "real_time": 1.4351768195629120e+08, + "cpu_time": 2.7431292680000752e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7407997724176307e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5, + "real_time": 1.4346329271793365e+08, + "cpu_time": 2.7423232179999101e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7422179696903634e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5, + "real_time": 1.4338159263134003e+08, + "cpu_time": 2.7410789240000212e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7443503182475595e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5, + "real_time": 1.4365660548210144e+08, + "cpu_time": 2.7442963779999447e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7371822214390984e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4345802724361420e+08, + "cpu_time": 2.7423605391999900e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7423583732340951e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time_median", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4346329271793365e+08, + "cpu_time": 2.7423232179999101e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.7422179696903634e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4478850690995081e+05, + "cpu_time": 1.4055532752561572e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7767410899875038e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0092743479880512e-03, + "cpu_time": 5.1253409431941050e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0091874463438134e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 2.9387553036212921e+08, + "cpu_time": 5.2167253650000590e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6537299402807631e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 2.9346024990081787e+08, + "cpu_time": 5.2176620200000912e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6589003940496116e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 2.9387488961219788e+08, + "cpu_time": 5.2239535100000012e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6537379066885481e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 2.9354144632816315e+08, + "cpu_time": 5.2173495799999613e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6578883065106106e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 2.9386170208454132e+08, + "cpu_time": 5.2220424300000447e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6539018741921473e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time_mean", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9372276365756989e+08, + "cpu_time": 5.2195465810000312e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6556316843443360e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time_median", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9386170208454132e+08, + "cpu_time": 5.2176620200000918e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6539018741921473e+09, + "dst_gpu": 1.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time_stddev", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0467859256882238e+05, + "cpu_time": 3.2399046548270422e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5481402139128842e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time_cv", + "family_index": 1, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/1/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.9684279835880315e-04, + "cpu_time": 6.2072530717913383e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9704511666905298e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2508, + "real_time": 2.6446617107648635e+05, + "cpu_time": 5.3297925996809243e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8719507898945929e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2508, + "real_time": 2.6436333136866614e+05, + "cpu_time": 5.3080345374801650e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8734570134917372e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2508, + "real_time": 2.6364243244737433e+05, + "cpu_time": 5.3046309848484863e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8840485216824897e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2508, + "real_time": 2.6403133953016170e+05, + "cpu_time": 5.3108141586922947e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8783274812080525e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2508, + "real_time": 2.6361990957552468e+05, + "cpu_time": 5.3036017783093464e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8843803628065255e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6402463679964264e+05, + "cpu_time": 5.3113748118022433e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8784328338166797e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time_median", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6403133953016170e+05, + "cpu_time": 5.3080345374801639e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8783274812080525e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9357202266354432e+02, + "cpu_time": 1.0683935679848159e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.7813456860746992e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4906640055799406e-03, + "cpu_time": 2.0115198151912975e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4906396304368652e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2645, + "real_time": 2.6390918738266808e+05, + "cpu_time": 5.3066829640832043e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7602451824854501e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2645, + "real_time": 2.6334740113887744e+05, + "cpu_time": 5.3061670737238904e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7767997373172399e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2645, + "real_time": 2.6401251082225848e+05, + "cpu_time": 5.3045704574669059e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7572081475289566e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2645, + "real_time": 2.6380272295206279e+05, + "cpu_time": 5.2990737542532699e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7633770306918127e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2645, + "real_time": 2.6380332587999845e+05, + "cpu_time": 5.3043951795841521e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7633592873336822e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6377502963517309e+05, + "cpu_time": 5.3041778858222859e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7641978770714290e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time_median", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6380332587999845e+05, + "cpu_time": 5.3045704574669059e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7633592873336822e+06, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5437636656194863e+02, + "cpu_time": 3.0201727494566001e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.4943981866249605e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.6436863987382087e-04, + "cpu_time": 5.6939507204864330e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6525079670583636e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2643, + "real_time": 2.6391868260109425e+05, + "cpu_time": 5.3076382633371919e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5519931971587589e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2643, + "real_time": 2.6329085593771806e+05, + "cpu_time": 5.3064584184638958e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5556939816280274e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2643, + "real_time": 2.6378719843243965e+05, + "cpu_time": 5.3064744797579315e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5527667848707430e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2643, + "real_time": 2.6396118458857283e+05, + "cpu_time": 5.3005160385925299e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5517433013434509e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2643, + "real_time": 2.6401106691323756e+05, + "cpu_time": 5.3071329360575252e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5514501145310231e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6379379769461253e+05, + "cpu_time": 5.3056440272418153e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5527294759064008e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time_median", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6391868260109430e+05, + "cpu_time": 5.3064744797579315e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5519931971587589e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9317848088350058e+02, + "cpu_time": 2.9087359418247775e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7277947160642016e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1113926235024904e-03, + "cpu_time": 5.4823428162347127e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1127467745504136e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2628, + "real_time": 2.6542689954515413e+05, + "cpu_time": 5.3387537823438540e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0863488267534792e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2628, + "real_time": 2.6515109580011235e+05, + "cpu_time": 5.3441253881279170e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0895591720185257e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2628, + "real_time": 2.6554646742756263e+05, + "cpu_time": 5.3341036035007879e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0849591332766131e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2628, + "real_time": 2.6653222953656170e+05, + "cpu_time": 5.3460307191780698e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0735494969009962e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2628, + "real_time": 2.6572072124128166e+05, + "cpu_time": 5.3337296270927927e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0829360848232236e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6567548271013453e+05, + "cpu_time": 5.3393486240486836e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0834705427545682e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time_median", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6554646742756263e+05, + "cpu_time": 5.3387537823438528e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0849591332766131e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2189122757209032e+02, + "cpu_time": 5.6326076191114691e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0474090097536813e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9643936363578579e-03, + "cpu_time": 1.0549241144775473e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9612345653710469e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2598, + "real_time": 2.6849602398661908e+05, + "cpu_time": 5.3984277559659712e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1021387790891543e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2598, + "real_time": 2.6766446256535838e+05, + "cpu_time": 5.3915292263279785e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1210964813826755e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2598, + "real_time": 2.6911130723732745e+05, + "cpu_time": 5.3940242571207497e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.0881871401825055e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2598, + "real_time": 2.6833686979058798e+05, + "cpu_time": 5.3871212856043235e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1057580394323707e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2598, + "real_time": 2.6891546387659502e+05, + "cpu_time": 5.3946807544265094e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.0926209909291774e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6850482549129758e+05, + "cpu_time": 5.3931566558891069e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1019602862031773e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time_median", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6849602398661908e+05, + "cpu_time": 5.3940242571207508e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1021387790891543e+07, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.6367996746909398e+02, + "cpu_time": 4.1813372774176423e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2821571108146224e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0993290025149406e-03, + "cpu_time": 7.7530425022084089e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1012216577574925e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2561, + "real_time": 2.7315948118136782e+05, + "cpu_time": 5.4772198945722496e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.1995922623034731e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2561, + "real_time": 2.7236921965678787e+05, + "cpu_time": 5.4688404490433482e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2030728010048610e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2561, + "real_time": 2.7252159630110627e+05, + "cpu_time": 5.4780187504880701e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2024001196512505e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2561, + "real_time": 2.7237196706520149e+05, + "cpu_time": 5.4702161577508890e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2030606656431666e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2561, + "real_time": 2.7257395183061325e+05, + "cpu_time": 5.4781334205389279e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2021691647323348e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7259924320701533e+05, + "cpu_time": 5.4744857344786974e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2020590026670173e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time_median", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7252159630110627e+05, + "cpu_time": 5.4772198945722496e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2024001196512505e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2599868172879167e+02, + "cpu_time": 4.5651391042132070e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4356549277246257e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1958899001096057e-03, + "cpu_time": 8.3389368894719003e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1943298328445837e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2496, + "real_time": 2.8156096507304121e+05, + "cpu_time": 5.6534947996794560e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3275953746997195e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2496, + "real_time": 2.8081404099840514e+05, + "cpu_time": 5.6514607772436179e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3337864362833697e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2496, + "real_time": 2.8178288869476557e+05, + "cpu_time": 5.6623375080127723e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3257622314671588e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2496, + "real_time": 2.8125128374785371e+05, + "cpu_time": 5.6588814903845696e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3301582530287781e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2496, + "real_time": 2.8111468016224814e+05, + "cpu_time": 5.6574100000002130e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3312905595031622e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8130477173526277e+05, + "cpu_time": 5.6567169150641258e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3297185709964377e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time_median", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8125128374785371e+05, + "cpu_time": 5.6574100000002119e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3301582530287781e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7882821973069571e+02, + "cpu_time": 4.3233618408881819e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1373787281742063e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3466825229940028e-03, + "cpu_time": 7.6428817383009735e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3466771339820357e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2368, + "real_time": 2.9536499770885613e+05, + "cpu_time": 5.9902065371621295e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4376280539916521e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2368, + "real_time": 2.9857744835385960e+05, + "cpu_time": 5.9982895270270691e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.3898827832656598e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2368, + "real_time": 2.9543540351270977e+05, + "cpu_time": 5.9849601097972703e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4365705139452326e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2368, + "real_time": 2.9533729264437931e+05, + "cpu_time": 5.9956708572633949e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4380443399617004e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2368, + "real_time": 2.9763021619736397e+05, + "cpu_time": 5.9866659459458827e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4038539391136205e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9646907168343384e+05, + "cpu_time": 5.9911585954391491e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4211959260555738e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time_median", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9543540351270983e+05, + "cpu_time": 5.9902065371621295e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4365705139452326e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5298596167855867e+03, + "cpu_time": 5.7166911502904907e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2756972859411682e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.1602671674944658e-03, + "cpu_time": 9.5418791861767748e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1472436960545665e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2123, + "real_time": 3.2931309474897111e+05, + "cpu_time": 6.6220964861046243e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9603272441937733e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2123, + "real_time": 3.2959487666457094e+05, + "cpu_time": 6.6272978944889258e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9535216885905743e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2123, + "real_time": 3.2940552669122827e+05, + "cpu_time": 6.6276686858218850e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9580935582092845e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2123, + "real_time": 3.3032577566276520e+05, + "cpu_time": 6.6244869477153884e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9359232404445159e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2123, + "real_time": 3.2940781163665879e+05, + "cpu_time": 6.6203308148846845e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9580383566965413e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2960941708083882e+05, + "cpu_time": 6.6243761658031028e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9531808176269388e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time_median", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2940781163665879e+05, + "cpu_time": 6.6244869477153884e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9580383566965413e+08, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1334126700439521e+02, + "cpu_time": 3.1996586851996511e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.9591952694984351e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2540335487533129e-03, + "cpu_time": 4.8301283096168239e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2522279447520531e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1773, + "real_time": 3.9401212971772027e+05, + "cpu_time": 7.9171547659335448e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3306392378722262e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1773, + "real_time": 3.9637197636764392e+05, + "cpu_time": 7.9585841060348286e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3227171224479077e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1773, + "real_time": 3.9443446800821333e+05, + "cpu_time": 7.9172670332769619e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3292144640591674e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1773, + "real_time": 3.9420498354997102e+05, + "cpu_time": 7.9140996446699055e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3299882596069188e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1773, + "real_time": 3.9439169759021676e+05, + "cpu_time": 7.9193275352510018e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3293586127787833e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9468305104675301e+05, + "cpu_time": 7.9252866170332499e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3283835393530006e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time_median", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9439169759021676e+05, + "cpu_time": 7.9172670332769607e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3293586127787833e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.5888386475494951e+02, + "cpu_time": 1.8707100965132311e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2176384002183960e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4295035274807457e-03, + "cpu_time": 2.3604321041117283e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4222209210643872e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1283, + "real_time": 5.3745240357066400e+05, + "cpu_time": 1.0835820194855507e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9510118347849824e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1283, + "real_time": 5.3621080549054919e+05, + "cpu_time": 1.0755401917380821e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9555294098199248e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1283, + "real_time": 5.3197309369002108e+05, + "cpu_time": 1.0708896500389830e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9711072090630991e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1283, + "real_time": 5.3448908358365181e+05, + "cpu_time": 1.0710240303975213e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9618286550765250e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1283, + "real_time": 5.3235119268942741e+05, + "cpu_time": 1.0721307365549554e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9697072428871913e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3449531580486277e+05, + "cpu_time": 1.0746333256430184e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9618368703263445e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time_median", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3448908358365181e+05, + "cpu_time": 1.0721307365549556e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9618286550765250e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3794001390804879e+03, + "cpu_time": 5.3433743772748294e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7299407865299974e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.4516763173078509e-03, + "cpu_time": 4.9722768220290987e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4498810877572114e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 894, + "real_time": 7.7588147779622464e+05, + "cpu_time": 1.5552310178971207e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7029282951265273e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 894, + "real_time": 7.7561783999927610e+05, + "cpu_time": 1.5549561935122577e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7038470388999271e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 894, + "real_time": 7.7550491010530607e+05, + "cpu_time": 1.5541468064877286e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7042407761354175e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 894, + "real_time": 7.7587019208604074e+05, + "cpu_time": 1.5551647908277144e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7029676115813904e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 894, + "real_time": 7.7574938918272196e+05, + "cpu_time": 1.5552472975391988e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7033885288771157e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7572476183391397e+05, + "cpu_time": 1.5549492212528042e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7034744501240759e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time_median", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7574938918272196e+05, + "cpu_time": 1.5551647908277144e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7033885288771157e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6285224731558947e+02, + "cpu_time": 4.6331607187891814e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6758941892885917e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0993560516308079e-04, + "cpu_time": 2.9796218779776608e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0994813503889768e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 521, + "real_time": 1.3373407558493330e+06, + "cpu_time": 2.6966973243761472e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1363016356562285e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 521, + "real_time": 1.3372940940857506e+06, + "cpu_time": 2.6967227984643108e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1364110695990639e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 521, + "real_time": 1.3376174962966775e+06, + "cpu_time": 2.6981923358924426e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1356527644205709e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 521, + "real_time": 1.3369615140573019e+06, + "cpu_time": 2.6964845067178239e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1371912773101954e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 521, + "real_time": 1.3372658272836923e+06, + "cpu_time": 2.6981708003837415e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1364773662986946e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3372959375145510e+06, + "cpu_time": 2.6972535531668933e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1364068226569510e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time_median", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3372940940857506e+06, + "cpu_time": 2.6967227984643104e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1364110695990639e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3353683218444786e+02, + "cpu_time": 8.5222927181819455e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4772978564251913e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7463362120017414e-04, + "cpu_time": 3.1596186825579568e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7463607771982833e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 278, + "real_time": 2.5055975696782200e+06, + "cpu_time": 5.0455149064746099e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3479470532362075e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 278, + "real_time": 2.5416980087154098e+06, + "cpu_time": 5.1012320431653233e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3003952362694955e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 278, + "real_time": 2.5347614863539021e+06, + "cpu_time": 5.1014719172663754e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3094269599568892e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 278, + "real_time": 2.5435719444758170e+06, + "cpu_time": 5.1035070287770294e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.2979637231093678e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 278, + "real_time": 2.5347321697932887e+06, + "cpu_time": 5.1022885935252095e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3094652365910926e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5320722358033275e+06, + "cpu_time": 5.0908028978417087e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3130396418326111e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time_median", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5347614863539017e+06, + "cpu_time": 5.1014719172663745e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3094269599568892e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5330639806435589e+04, + "cpu_time": 2.5332357558479718e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0196036315625895e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.0545823257572972e-03, + "cpu_time": 4.9761026044083528e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0959235321598618e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 139, + "real_time": 4.9914713151645744e+06, + "cpu_time": 9.9506470431653876e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3611764829799156e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 139, + "real_time": 4.9975973674558476e+06, + "cpu_time": 9.9559767338133901e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3570563545700088e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 139, + "real_time": 4.9957268965169247e+06, + "cpu_time": 9.9538009208633006e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3583132840382562e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 139, + "real_time": 4.9995853089921763e+06, + "cpu_time": 9.9608261151084173e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3557215175076141e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 139, + "real_time": 4.9928330360717485e+06, + "cpu_time": 9.9506875539569091e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3602597721152611e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9954427848402541e+06, + "cpu_time": 9.9543876733814813e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3585054822422113e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time_median", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9957268965169247e+06, + "cpu_time": 9.9538009208632987e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3583132840382562e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3341720126328560e+03, + "cpu_time": 4.2428210409526209e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2416037053502565e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.6744273855985669e-04, + "cpu_time": 4.2622622105608076e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6744083557479061e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 69, + "real_time": 9.8551584920589477e+06, + "cpu_time": 1.9677039246376824e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4047582316446118e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 69, + "real_time": 9.8767284492867589e+06, + "cpu_time": 1.9755814637680512e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3973225215504546e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 69, + "real_time": 9.8508546749750767e+06, + "cpu_time": 1.9728923333332203e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4062457631459165e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 69, + "real_time": 9.8941825681190565e+06, + "cpu_time": 1.9738983507245988e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3913293765286665e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 69, + "real_time": 9.8775355415283777e+06, + "cpu_time": 1.9697673565217327e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3970449267356453e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8708919451936446e+06, + "cpu_time": 1.9719686857970573e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3993401639210587e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time_median", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8767284492867589e+06, + "cpu_time": 1.9728923333332203e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3973225215504546e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7816347927917799e+04, + "cpu_time": 3.1883747622511357e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1350319405329917e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8049379961648730e-03, + "cpu_time": 1.6168485763567867e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8047714099480932e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1335860432097405e+07, + "cpu_time": 4.2675718484846666e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1453554082610254e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1348320625045083e+07, + "cpu_time": 4.2700432666668519e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1435195853893204e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1277512959910162e+07, + "cpu_time": 4.2515126757576361e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1539806426833143e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1340655275818072e+07, + "cpu_time": 4.2700585090907834e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1446487060799704e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1284237504005432e+07, + "cpu_time": 4.2576889545454256e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1529841737281375e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1317317359375235e+07, + "cpu_time": 4.2633750509090729e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1480977032283540e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time_median", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1335860432097405e+07, + "cpu_time": 4.2675718484846666e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1453554082610254e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3646594996874803e+04, + "cpu_time": 8.3635288700982492e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9714559370228760e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5783690991529580e-03, + "cpu_time": 1.9617154883699258e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5791936609606111e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18, + "real_time": 3.9461461206277214e+07, + "cpu_time": 7.7981712388887241e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4012356333791747e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18, + "real_time": 3.9421541823281184e+07, + "cpu_time": 7.7887997277779654e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4046798220544238e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18, + "real_time": 3.9433871499366231e+07, + "cpu_time": 7.7994282055556133e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4036152905290341e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18, + "real_time": 3.9458120448721781e+07, + "cpu_time": 7.8052957611109570e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4015236020789199e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18, + "real_time": 3.9418208516306348e+07, + "cpu_time": 7.7941505944446981e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4049677306993141e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9438640698790550e+07, + "cpu_time": 7.7971691055555940e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4032044157481737e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time_median", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9433871499366231e+07, + "cpu_time": 7.7981712388887256e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4036152905290341e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0204127718147698e+04, + "cpu_time": 6.1503704791058684e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7432842619377943e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.1229269975238496e-04, + "cpu_time": 7.8879531735737802e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1224788433830936e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9, + "real_time": 7.5653250018755600e+07, + "cpu_time": 1.4766938899999282e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5482342917647390e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9, + "real_time": 7.5720396306779653e+07, + "cpu_time": 1.4773872899999988e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5450878375284672e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9, + "real_time": 7.5604698724216878e+07, + "cpu_time": 1.4759300688889551e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5505128719469075e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9, + "real_time": 7.5640397767225906e+07, + "cpu_time": 1.4768474888887796e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5488371812384882e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9, + "real_time": 7.5695843332343638e+07, + "cpu_time": 1.4771011322222370e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5462377348968883e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5662917229864344e+07, + "cpu_time": 1.4767919739999801e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5477819834750986e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time_median", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5653250018755600e+07, + "cpu_time": 1.4768474888887793e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5482342917647390e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5765565843232711e+04, + "cpu_time": 5.4899060498109844e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1458624716416476e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.0486123875182706e-04, + "cpu_time": 3.7174538773671989e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0484620578058954e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5, + "real_time": 1.4917475581169128e+08, + "cpu_time": 2.8526693339999837e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5989394390409570e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5, + "real_time": 1.4938601851463318e+08, + "cpu_time": 2.8542624419999355e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5938497949017267e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5, + "real_time": 1.4938009679317474e+08, + "cpu_time": 2.8540177840000069e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5939922621909156e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5, + "real_time": 1.4939443767070770e+08, + "cpu_time": 2.8491675279999530e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5936472627137589e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5, + "real_time": 1.4956922233104706e+08, + "cpu_time": 2.8567448359999621e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5894477729631023e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4938090622425079e+08, + "cpu_time": 2.8533723847999680e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5939753063620930e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time_median", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4938601851463318e+08, + "cpu_time": 2.8540177840000069e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5938497949017267e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3979524258852372e+05, + "cpu_time": 2.7729618722427828e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3639569397481889e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.3583073045937302e-04, + "cpu_time": 9.7181913128986056e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.3599890177132738e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 2.9438081383705139e+08, + "cpu_time": 5.2322199650001264e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6474585758647585e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 2.9471392929553986e+08, + "cpu_time": 5.2382326349999177e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6433358496715269e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 2.9458577930927277e+08, + "cpu_time": 5.2325608550000882e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6449207647349644e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 2.9473392665386200e+08, + "cpu_time": 5.2387826699998641e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6430886535197268e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 2.9491466283798218e+08, + "cpu_time": 5.2407752499999559e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6408560146427293e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time_mean", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9466582238674164e+08, + "cpu_time": 5.2365142749999905e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6439319716867418e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time_median", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9471392929553986e+08, + "cpu_time": 5.2382326349999177e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6433358496715269e+09, + "dst_gpu": 2.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time_stddev", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9780723172584400e+05, + "cpu_time": 3.8834557092098950e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4465642969029038e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time_cv", + "family_index": 2, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/2/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.7129343377402906e-04, + "cpu_time": 7.4161083218089805e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7140778585128536e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2700, + "real_time": 2.5914309987220570e+05, + "cpu_time": 5.2335482740739250e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9514847221669313e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2700, + "real_time": 2.5848383904236404e+05, + "cpu_time": 5.2309596962960245e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9615629502940499e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2700, + "real_time": 2.5894950397304018e+05, + "cpu_time": 5.2144215629629832e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9544389322584337e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2700, + "real_time": 2.5829415012763263e+05, + "cpu_time": 5.2226520629628957e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9644722867087931e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2700, + "real_time": 2.5846772655818821e+05, + "cpu_time": 5.2174145037034346e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9618099080910571e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5866766391468616e+05, + "cpu_time": 5.2237992199998524e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9587537599038528e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time_median", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5848383904236401e+05, + "cpu_time": 5.2226520629628963e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9615629502940499e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6012228740007333e+02, + "cpu_time": 8.3117010755999638e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5088517053753048e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3922199704051487e-03, + "cpu_time": 1.5911218493577929e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3915621024908353e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2705, + "real_time": 2.5889696510312406e+05, + "cpu_time": 5.2309683142328158e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9104828408638891e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2705, + "real_time": 2.5889880129122373e+05, + "cpu_time": 5.2302892680223071e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9104267373424256e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2705, + "real_time": 2.5887455563783963e+05, + "cpu_time": 5.2286938853974658e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9111676114863576e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2705, + "real_time": 2.5855313069634690e+05, + "cpu_time": 5.2281130573013856e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9210025207748767e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2705, + "real_time": 2.5833906711710966e+05, + "cpu_time": 5.2237181552680105e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9275659808417810e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5871250396912885e+05, + "cpu_time": 5.2283565360443975e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9161291382618658e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time_median", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5887455563783963e+05, + "cpu_time": 5.2286938853974658e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9111676114863576e+06, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5487657209930185e+02, + "cpu_time": 2.8392120792381195e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8033009666783009e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.8517299391804930e-04, + "cpu_time": 5.4304102248278849e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.8574705268030297e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2704, + "real_time": 2.5885460821258320e+05, + "cpu_time": 5.2290691013312386e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5823554497574091e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2704, + "real_time": 2.5823188318541003e+05, + "cpu_time": 5.2179243158283114e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5861712928217618e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2704, + "real_time": 2.5897555188162214e+05, + "cpu_time": 5.2249207655323972e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5816164770148974e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2704, + "real_time": 2.5826667911805675e+05, + "cpu_time": 5.2213778328405577e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5859575900333894e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2704, + "real_time": 2.5871419696551992e+05, + "cpu_time": 5.2312897041417728e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5832142371939074e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5860858387263841e+05, + "cpu_time": 5.2249163439348555e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5838630093642734e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time_median", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5871419696551995e+05, + "cpu_time": 5.2249207655323966e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5832142371939074e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4100949025589000e+02, + "cpu_time": 5.4606467489355941e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0890182854333754e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3186317528571790e-03, + "cpu_time": 1.0451165893353256e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3189387422286349e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2676, + "real_time": 2.6074049415361619e+05, + "cpu_time": 5.2619005530645186e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1418211530939471e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2676, + "real_time": 2.6146187396329912e+05, + "cpu_time": 5.2619774962631834e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1331527904331837e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2676, + "real_time": 2.6074796940068822e+05, + "cpu_time": 5.2585937892376666e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1417310818675842e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2676, + "real_time": 2.6090928507023543e+05, + "cpu_time": 5.2621406390131032e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1397886042249348e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2676, + "real_time": 2.6076602549976020e+05, + "cpu_time": 5.2535176233182533e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1415135404621691e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6092512961751985e+05, + "cpu_time": 5.2596260201793443e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1396014340163637e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time_median", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6076602549976017e+05, + "cpu_time": 5.2619005530645186e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1415135404621691e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0786996428838654e+02, + "cpu_time": 3.7217019104855927e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6992817060376736e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1799168778403267e-03, + "cpu_time": 7.0759820112812678e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1782647523209129e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2657, + "real_time": 2.6347265469664295e+05, + "cpu_time": 5.3119289461801725e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2184821490731940e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2657, + "real_time": 2.6417914609527122e+05, + "cpu_time": 5.3152410726383619e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2018521303310670e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2657, + "real_time": 2.6367553573522024e+05, + "cpu_time": 5.3169307263834076e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2136974347338051e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2657, + "real_time": 2.6375646663073968e+05, + "cpu_time": 5.3176634362062346e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2117908270805269e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2657, + "real_time": 2.6420678520453535e+05, + "cpu_time": 5.3144677982684504e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2012033443109140e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6385811767248192e+05, + "cpu_time": 5.3152463959353254e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2094051771059014e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time_median", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6375646663073974e+05, + "cpu_time": 5.3152410726383608e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2117908270805269e+07, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2283087676397236e+02, + "cpu_time": 2.2522637662708502e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5964596729331752e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2235017804708638e-03, + "cpu_time": 4.2373647400301167e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2233796082338690e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2619, + "real_time": 2.6763668506516126e+05, + "cpu_time": 5.4008392859870160e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2243463556583060e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2619, + "real_time": 2.6838060158572282e+05, + "cpu_time": 5.4000471744940057e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2209526249807458e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2619, + "real_time": 2.6773288891577482e+05, + "cpu_time": 5.3989539251623326e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2239064140643689e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2619, + "real_time": 2.6771780530834635e+05, + "cpu_time": 5.3983248377244454e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2239753707176541e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2619, + "real_time": 2.6685805530920299e+05, + "cpu_time": 5.3827396449025464e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2279187136409424e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6766520723684167e+05, + "cpu_time": 5.3961809736540692e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2242198958124036e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time_median", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6771780530834640e+05, + "cpu_time": 5.3989539251623326e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2239753707176541e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.4116637683812371e+02, + "cpu_time": 7.5763136570856136e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4764070463576863e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0218032161321451e-03, + "cpu_time": 1.4040140043626540e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0228449601485360e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2534, + "real_time": 2.7651654883220221e+05, + "cpu_time": 5.5771938397789316e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3700570644605085e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2534, + "real_time": 2.7645996771247481e+05, + "cpu_time": 5.5694181176005490e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3705421273925295e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2534, + "real_time": 2.7584206404685596e+05, + "cpu_time": 5.5700695264408237e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3758522916529411e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2534, + "real_time": 2.7596280079363642e+05, + "cpu_time": 5.5778070994475961e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3748128302628544e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2534, + "real_time": 2.7578144888814399e+05, + "cpu_time": 5.5758506945541641e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3763744901703367e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7611256605466269e+05, + "cpu_time": 5.5740678555644129e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3735277607878339e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time_median", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7596280079363647e+05, + "cpu_time": 5.5758506945541641e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3748128302628544e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4968898466715228e+02, + "cpu_time": 4.0168064395597651e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0049315819166333e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2664725465552462e-03, + "cpu_time": 7.2062388611755335e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2660191431336874e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2357, + "real_time": 2.9344090206876927e+05, + "cpu_time": 5.9024817225287796e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4667256362673891e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2357, + "real_time": 2.9434137738605752e+05, + "cpu_time": 5.9196606788289861e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4530606319778901e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2357, + "real_time": 2.9723320523955364e+05, + "cpu_time": 5.9148471616461151e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4097361159350675e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2357, + "real_time": 2.9356105611533148e+05, + "cpu_time": 5.9003363428088778e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4648974129765254e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2357, + "real_time": 2.9406122150625457e+05, + "cpu_time": 5.9109406618583994e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4573031196910864e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9452755246319331e+05, + "cpu_time": 5.9096533135342319e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4503445833695918e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time_median", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9406122150625457e+05, + "cpu_time": 5.9109406618583994e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4573031196910864e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5562460391138391e+03, + "cpu_time": 8.1703479898862247e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3371828762336932e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.2838725141286097e-03, + "cpu_time": 1.3825426901396350e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2516896893051099e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2152, + "real_time": 3.2548742122083926e+05, + "cpu_time": 6.5495666310408060e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0538903474902177e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2152, + "real_time": 3.2558370108731772e+05, + "cpu_time": 6.5444220353161462e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0515086942173457e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2152, + "real_time": 3.2519872474679322e+05, + "cpu_time": 6.5489898745350866e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0610402209944391e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2152, + "real_time": 3.2557373765062325e+05, + "cpu_time": 6.5508490985134954e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0517550921539497e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2152, + "real_time": 3.2648921025482868e+05, + "cpu_time": 6.5554890241636336e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0291780483463311e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2566655899208051e+05, + "cpu_time": 6.5498633327138331e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0494744806404579e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time_median", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2557373765062325e+05, + "cpu_time": 6.5495666310408048e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0517550921539497e+08, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8559127774655661e+02, + "cpu_time": 3.9711086923042524e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1984892651667765e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4910688995807065e-03, + "cpu_time": 6.0628878658738146e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4889037390567869e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1793, + "real_time": 3.9067731910383637e+05, + "cpu_time": 7.8485188455104001e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3419975370022745e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1793, + "real_time": 3.8991961099038069e+05, + "cpu_time": 7.8368048689348483e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3446053628037040e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1793, + "real_time": 3.8977746299662249e+05, + "cpu_time": 7.8322264528722956e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3450957271086326e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1793, + "real_time": 3.8996985316176823e+05, + "cpu_time": 7.8320678750696662e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3444321291741328e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1793, + "real_time": 3.8991471408018132e+05, + "cpu_time": 7.8338256999444612e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3446222496035028e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9005179206655780e+05, + "cpu_time": 7.8366887484663352e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3441506011384497e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time_median", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8991961099038069e+05, + "cpu_time": 7.8338256999444612e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3446053628037040e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5690219545736988e+02, + "cpu_time": 6.8816109993412692e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2285628547209133e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.1501232071372889e-04, + "cpu_time": 8.7812738520309656e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.1400684839954883e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1286, + "real_time": 5.3203970418908400e+05, + "cpu_time": 1.0667382791601680e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9708604296707561e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1286, + "real_time": 5.3484656196775590e+05, + "cpu_time": 1.0670331018662665e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9605174167001846e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1286, + "real_time": 5.3309298421525653e+05, + "cpu_time": 1.0679353429237213e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9669664224592342e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1286, + "real_time": 5.3422035240969423e+05, + "cpu_time": 1.0659408328148904e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9628155222282615e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1286, + "real_time": 5.3249854372863541e+05, + "cpu_time": 1.0673157037324575e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9691621927408705e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3333962930208526e+05, + "cpu_time": 1.0669926520995009e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9660643967598619e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time_median", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3309298421525653e+05, + "cpu_time": 1.0670331018662665e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9669664224592342e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1730924186785892e+03, + "cpu_time": 7.3581484249752680e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3225020381070962e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1995223197902401e-03, + "cpu_time": 6.8961566047308592e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1985556756079403e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 892, + "real_time": 7.8526397617826145e+05, + "cpu_time": 1.5736686524663193e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.6706331419995370e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 892, + "real_time": 7.8426828069430706e+05, + "cpu_time": 1.5720164360986049e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.6740237385903282e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 892, + "real_time": 7.8434541458921775e+05, + "cpu_time": 1.5718973968609849e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.6737607704359865e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 892, + "real_time": 7.8462163669489441e+05, + "cpu_time": 1.5724869080718532e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.6728194863883061e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 892, + "real_time": 7.8537879013686103e+05, + "cpu_time": 1.5731990067264389e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.6702427240676413e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8477561965870834e+05, + "cpu_time": 1.5726536800448406e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.6722959722963600e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time_median", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8462163669489429e+05, + "cpu_time": 1.5724869080718532e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.6728194863883061e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1683815046706991e+02, + "cpu_time": 7.6377470472420123e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7596668608347434e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.5858079369468438e-04, + "cpu_time": 4.8565982098641323e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5848501778140417e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 522, + "real_time": 1.3328127173164510e+06, + "cpu_time": 2.6859470402298719e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1469567670730309e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 522, + "real_time": 1.3324801608923664e+06, + "cpu_time": 2.6866224904213916e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1477421751563339e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 522, + "real_time": 1.3323164914317173e+06, + "cpu_time": 2.6857274463601289e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1481288620039296e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 522, + "real_time": 1.3320479728013875e+06, + "cpu_time": 2.6863781072795545e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1487634722187166e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 522, + "real_time": 1.3323342585954031e+06, + "cpu_time": 2.6872896436782368e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1480868805563803e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3323983202074650e+06, + "cpu_time": 2.6863929455938367e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1479356314016786e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time_median", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3323342585954028e+06, + "cpu_time": 2.6863781072795545e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1480868805563803e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7926958577383056e+02, + "cpu_time": 6.1210686871404073e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.5976046271355182e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0959917281368687e-04, + "cpu_time": 2.2785455482899667e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0958511861939846e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 277, + "real_time": 2.5184852792293048e+06, + "cpu_time": 5.0569727689530449e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3308147834666095e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 277, + "real_time": 2.5246826022365787e+06, + "cpu_time": 5.0616007509025401e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3226386527037725e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 277, + "real_time": 2.5194781884510703e+06, + "cpu_time": 5.0579627075812463e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3295021320097895e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 277, + "real_time": 2.5228601779506309e+06, + "cpu_time": 5.0597470758126676e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3250388084583554e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 277, + "real_time": 2.5199888289203388e+06, + "cpu_time": 5.0583555379061075e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3288274549986815e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5210990153575847e+06, + "cpu_time": 5.0589277682311218e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3273643663274422e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time_median", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5199888289203392e+06, + "cpu_time": 5.0583555379061075e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3288274549986815e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5804553344245492e+03, + "cpu_time": 1.7956136163185595e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4042854483371396e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0235438269998076e-03, + "cpu_time": 3.5493956399112699e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0231177212775763e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 139, + "real_time": 4.9661736325662341e+06, + "cpu_time": 9.8941803669064194e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3782983120005202e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 139, + "real_time": 4.9629424323441740e+06, + "cpu_time": 9.8904412086335644e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3804978052255030e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 139, + "real_time": 4.9611536832861332e+06, + "cpu_time": 9.8893235611517765e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3817166471826019e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 139, + "real_time": 4.9632509346220559e+06, + "cpu_time": 9.8926301942444406e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3802876826088905e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 139, + "real_time": 4.9636135373612959e+06, + "cpu_time": 9.8897967482010331e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3800407452589321e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9634268440359784e+06, + "cpu_time": 9.8912744158274475e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3801682384552898e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time_median", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9632509346220559e+06, + "cpu_time": 9.8904412086335625e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3802876826088905e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8036463241908982e+03, + "cpu_time": 2.0595057968142837e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2281419102367610e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.6338730898353983e-04, + "cpu_time": 2.0821440294071524e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.6333750973235938e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 70, + "real_time": 9.8072688095271587e+06, + "cpu_time": 1.9577797800000750e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4213839399817338e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 70, + "real_time": 9.8084711735802032e+06, + "cpu_time": 1.9568157300000168e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4209645322077498e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 70, + "real_time": 9.8057991159813739e+06, + "cpu_time": 1.9553880385714300e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4218967371372504e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 70, + "real_time": 9.8007020141397212e+06, + "cpu_time": 1.9545879814285982e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4236763806909108e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 70, + "real_time": 9.8025008769971989e+06, + "cpu_time": 1.9562303328572333e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4230480997701001e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8049483980451338e+06, + "cpu_time": 1.9561603725714706e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4221939379575491e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time_median", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8057991159813739e+06, + "cpu_time": 1.9562303328572329e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4218967371372504e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2611391156068466e+03, + "cpu_time": 1.2375158520325886e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1383202985451855e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.3260135425669746e-04, + "cpu_time": 6.3262494700565481e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3262881040124905e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1265816564361256e+07, + "cpu_time": 4.2314153909091361e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1557153611710229e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1225191144780681e+07, + "cpu_time": 4.2187754454547837e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1617554603979249e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1283630563905746e+07, + "cpu_time": 4.2343828030301332e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1530740866085067e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1269821815870024e+07, + "cpu_time": 4.2339913121209480e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1551211185948043e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1245991303162143e+07, + "cpu_time": 4.2218214606061168e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1586600522618055e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1258090278415971e+07, + "cpu_time": 4.2280772824242242e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1568652158068132e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time_median", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1265816564361256e+07, + "cpu_time": 4.2314153909091368e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1557153611710229e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2794565384289559e+04, + "cpu_time": 7.2721692643859147e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3864668693610453e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0722771935649186e-03, + "cpu_time": 1.7199707523359933e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0727309016566777e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18, + "real_time": 3.9377885146273509e+07, + "cpu_time": 7.7484670388887227e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4084544535958042e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18, + "real_time": 3.9347022357914180e+07, + "cpu_time": 7.7484118833334982e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4111279572596097e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18, + "real_time": 3.9294177252385348e+07, + "cpu_time": 7.7485219777776778e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4157154414488297e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18, + "real_time": 3.9349724435144000e+07, + "cpu_time": 7.7533814277779177e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4108937210276256e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18, + "real_time": 3.9344818227820925e+07, + "cpu_time": 7.7566624666663602e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4113190515414286e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9342725483907595e+07, + "cpu_time": 7.7510889588888377e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4115021249746599e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time_median", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9347022357914187e+07, + "cpu_time": 7.7485219777776763e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4111279572596097e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.0269459278526956e+04, + "cpu_time": 3.7732609425167619e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6260240505829342e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.6937880907381742e-04, + "cpu_time": 4.8680397845126533e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6975594749261367e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9, + "real_time": 7.4135813448164195e+07, + "cpu_time": 1.4447255944444096e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6208607353811560e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9, + "real_time": 7.4383102357387543e+07, + "cpu_time": 1.4493691677777559e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6088230726146860e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9, + "real_time": 7.4213820199171707e+07, + "cpu_time": 1.4459334366665643e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6170548191642075e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9, + "real_time": 7.4103632734881505e+07, + "cpu_time": 1.4441456622221771e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6224331533161678e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9, + "real_time": 7.4176327221923411e+07, + "cpu_time": 1.4449403644444677e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6188830864715796e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4202539192305669e+07, + "cpu_time": 1.4458228451110750e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6176109733895593e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time_median", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4176327221923411e+07, + "cpu_time": 1.4449403644444677e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.6188830864715796e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0914691628988137e+05, + "cpu_time": 2.0847247146067693e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3143473611121811e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4709323626650123e-03, + "cpu_time": 1.4418949884877567e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4690212408695915e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5, + "real_time": 1.4928273856639862e+08, + "cpu_time": 2.8460493640000093e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.5963361682382874e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5, + "real_time": 1.4890683293342590e+08, + "cpu_time": 2.8406827800001794e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6054148854272337e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5, + "real_time": 1.4841701090335846e+08, + "cpu_time": 2.8378401920001578e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6173138694295816e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5, + "real_time": 1.4853525161743164e+08, + "cpu_time": 2.8366371540000731e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6144343255483098e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5, + "real_time": 1.4880872070789337e+08, + "cpu_time": 2.8359524320001131e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6077919993268404e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4879011094570160e+08, + "cpu_time": 2.8394323844001067e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6082582495940509e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time_median", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4880872070789337e+08, + "cpu_time": 2.8378401920001578e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6077919993268404e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3941465686100366e+05, + "cpu_time": 4.1172310632484750e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2247537615663605e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2811640820999670e-03, + "cpu_time": 1.4500190551705395e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2794249171304995e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 2.9239414632320404e+08, + "cpu_time": 5.1977951649996614e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6722411768569293e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 2.9252967238426208e+08, + "cpu_time": 5.1995759999999791e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6705398643784442e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 2.9191213846206665e+08, + "cpu_time": 5.1884439349998957e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6783048134174471e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 2.9218535125255585e+08, + "cpu_time": 5.1905691700000036e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6748653530953074e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 2.9234902560710907e+08, + "cpu_time": 5.2007121849999297e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6728079451272497e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time_mean", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9227406680583954e+08, + "cpu_time": 5.1954192909998941e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6737518305750756e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time_median", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9234902560710907e+08, + "cpu_time": 5.1977951649996614e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6728079451272497e+09, + "dst_gpu": 3.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time_stddev", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.3676713216210334e+05, + "cpu_time": 5.5479049846219795e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9773858669913779e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time_cv", + "family_index": 3, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/3/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.1008600848391378e-04, + "cpu_time": 1.0678454757699156e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1044828401631824e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2672, + "real_time": 2.6210799488379454e+05, + "cpu_time": 5.2909412724548473e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9067865917405146e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2672, + "real_time": 2.6267171223682049e+05, + "cpu_time": 5.2876738547903509e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8984022728598136e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2672, + "real_time": 2.6273188433004473e+05, + "cpu_time": 5.2920086302393849e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8975094424156281e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2672, + "real_time": 2.6194643153698172e+05, + "cpu_time": 5.2882801422156719e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9091962199738203e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2672, + "real_time": 2.6253015074800857e+05, + "cpu_time": 5.2889341953594529e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9005043690501428e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6239763474712998e+05, + "cpu_time": 5.2895676190119411e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9024797792079840e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time_median", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6253015074800857e+05, + "cpu_time": 5.2889341953594517e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9005043690501428e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5066983467558299e+02, + "cpu_time": 1.8364792771678245e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2175607779106331e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3364062332860585e-03, + "cpu_time": 3.4718892156082641e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3369859866306719e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2672, + "real_time": 2.6238727516461228e+05, + "cpu_time": 5.2869500748502044e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8052565571831129e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2672, + "real_time": 2.6187301890007540e+05, + "cpu_time": 5.2808297492513852e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8205842228498869e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2672, + "real_time": 2.6251979161424754e+05, + "cpu_time": 5.2868573166169075e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8013165689594075e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2672, + "real_time": 2.6191266086592106e+05, + "cpu_time": 5.2882357335329533e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8194005330976220e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2672, + "real_time": 2.6251463823210070e+05, + "cpu_time": 5.2875108982032584e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8014697153355442e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6224147695539141e+05, + "cpu_time": 5.2860767544909415e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8096055194851151e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time_median", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6238727516461234e+05, + "cpu_time": 5.2869500748502044e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8052565571831129e+06, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2296075978248228e+02, + "cpu_time": 2.9841660630138290e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.6214045526838891e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2315395853167023e-03, + "cpu_time": 5.6453324490200469e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2319962293458101e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2674, + "real_time": 2.6259252490864461e+05, + "cpu_time": 5.2875673111442686e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5598311495824145e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2674, + "real_time": 2.6249708045307454e+05, + "cpu_time": 5.2833980067313428e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5603983072612589e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2674, + "real_time": 2.6203233973494411e+05, + "cpu_time": 5.2881074121163622e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5631658306540573e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2674, + "real_time": 2.6184751134276635e+05, + "cpu_time": 5.2831273223633715e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5642692111128036e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2674, + "real_time": 2.6256673642242269e+05, + "cpu_time": 5.2874738107706129e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5599843513346914e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6230723857237044e+05, + "cpu_time": 5.2859347726251918e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5615297699890450e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time_median", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6249708045307454e+05, + "cpu_time": 5.2874738107706117e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5603983072612589e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4339695104597570e+02, + "cpu_time": 2.4531174522640762e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0454260057312938e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3091402010670498e-03, + "cpu_time": 4.6408394310279514e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3098860137297566e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2673, + "real_time": 2.6415794921400957e+05, + "cpu_time": 5.3224138346426480e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1011748934207499e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2673, + "real_time": 2.6437278089695045e+05, + "cpu_time": 5.3231657201643160e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0986548510049339e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2673, + "real_time": 2.6347334763870336e+05, + "cpu_time": 5.3197379274222930e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1092328971481219e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2673, + "real_time": 2.6409066822435579e+05, + "cpu_time": 5.3193167227833893e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1019649634270921e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2673, + "real_time": 2.6421558938833739e+05, + "cpu_time": 5.3161616647961095e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1004983540012110e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6406206707247131e+05, + "cpu_time": 5.3201591739617509e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1023051918004222e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time_median", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6415794921400957e+05, + "cpu_time": 5.3197379274222930e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1011748934207499e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4522245175559840e+02, + "cpu_time": 2.7834361408645873e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0612831069749373e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3073534399806572e-03, + "cpu_time": 5.2318662841658030e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3091178513671546e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2642, + "real_time": 2.6712809677401814e+05, + "cpu_time": 5.3797061809238326e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1333870146427698e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2642, + "real_time": 2.6718968690149515e+05, + "cpu_time": 5.3818597918241704e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1319732022592217e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2642, + "real_time": 2.6662616702464648e+05, + "cpu_time": 5.3811820590460917e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1449332534887649e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2642, + "real_time": 2.6742393267208769e+05, + "cpu_time": 5.3829309538224945e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1266019971704938e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2642, + "real_time": 2.6710896226129780e+05, + "cpu_time": 5.3769828993188334e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1338263835462205e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6709536912670906e+05, + "cpu_time": 5.3805323769870843e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1341443702214934e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time_median", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6712809677401814e+05, + "cpu_time": 5.3811820590460917e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1333870146427698e+07, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9079904176498468e+02, + "cpu_time": 2.3031868888148787e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6834883626553885e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0887461011240172e-03, + "cpu_time": 4.2805929366130589e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0895551130326688e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2603, + "real_time": 2.7117865548576106e+05, + "cpu_time": 5.4593341029581102e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2083546893210617e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2603, + "real_time": 2.7142747287578380e+05, + "cpu_time": 5.4633009412217594e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2072469913535969e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2603, + "real_time": 2.7087875207393424e+05, + "cpu_time": 5.4605154821361834e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2096925192218927e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2603, + "real_time": 2.7075932143650349e+05, + "cpu_time": 5.4596611409912386e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2102261087873390e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2603, + "real_time": 2.7134670884061506e+05, + "cpu_time": 5.4626064809833269e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2076063181310751e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7111818214251951e+05, + "cpu_time": 5.4610836296581232e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2086253253629930e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time_median", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7117865548576100e+05, + "cpu_time": 5.4605154821361834e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2083546893210617e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9053829299088511e+02, + "cpu_time": 1.7778256378384444e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2954569224794779e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0716296881857852e-03, + "cpu_time": 3.2554448135227340e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0718432712721838e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2505, + "real_time": 2.8026273900256335e+05, + "cpu_time": 5.6345172335328721e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3383772039493483e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2505, + "real_time": 2.7922807176998944e+05, + "cpu_time": 5.6299891497008514e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3470419569413653e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2505, + "real_time": 2.8023687428320770e+05, + "cpu_time": 5.6409221157684142e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3385930266182327e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2505, + "real_time": 2.7935051128264522e+05, + "cpu_time": 5.6346426227545075e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3460132469094017e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2505, + "real_time": 2.8001389841729414e+05, + "cpu_time": 5.6328924231538281e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3404552549150318e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7981841895114002e+05, + "cpu_time": 5.6345927089820954e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3420961378666759e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time_median", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8001389841729414e+05, + "cpu_time": 5.6345172335328721e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3404552549150318e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9451612086650158e+02, + "cpu_time": 4.0047880133561773e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.1412445092025172e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7672750876090491e-03, + "cpu_time": 7.1075022103587917e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7681787020812116e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2353, + "real_time": 2.9915434803251264e+05, + "cpu_time": 5.9723498725032876e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.3814171801960528e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2353, + "real_time": 2.9694819546835410e+05, + "cpu_time": 5.9705999787500978e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4139685642227918e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2353, + "real_time": 2.9644881983630400e+05, + "cpu_time": 5.9681155376114929e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4214040073553550e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2353, + "real_time": 2.9904759228809370e+05, + "cpu_time": 5.9739783977899782e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.3829812839198202e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2353, + "real_time": 2.9658338249106111e+05, + "cpu_time": 5.9639090055248444e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4193979750011933e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9763646762326511e+05, + "cpu_time": 5.9697905584359402e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4038338021390432e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time_median", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9694819546835416e+05, + "cpu_time": 5.9705999787500966e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4139685642227918e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3498562772749333e+03, + "cpu_time": 3.9414974184637157e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9943679738904755e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.5352516378588418e-03, + "cpu_time": 6.6024048580631811e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5287085378239410e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2131, + "real_time": 3.2854220429852739e+05, + "cpu_time": 6.6106657015484560e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9790053323500812e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2131, + "real_time": 3.2835765109156666e+05, + "cpu_time": 6.6150664148288826e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9834899271738875e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2131, + "real_time": 3.2846915268709383e+05, + "cpu_time": 6.6150161379633681e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9807798648819697e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2131, + "real_time": 3.2897392792462453e+05, + "cpu_time": 6.6110278320038947e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9685342134487689e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2131, + "real_time": 3.2819028980815841e+05, + "cpu_time": 6.6101279727823054e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9875611235553205e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2850664516199421e+05, + "cpu_time": 6.6123808118253818e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9798740922820055e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time_median", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2846915268709383e+05, + "cpu_time": 6.6110278320038947e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9807798648819697e+08, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9296043290740977e+02, + "cpu_time": 2.4497371663926964e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.1121855217647413e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.9179454121223107e-04, + "cpu_time": 3.7047732671591764e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9126538082142455e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1780, + "real_time": 3.9259472406035021e+05, + "cpu_time": 7.8952351910109760e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3354433156350970e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1780, + "real_time": 3.9425793607795110e+05, + "cpu_time": 7.9301382134831382e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3298096297453854e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1780, + "real_time": 3.9469622825390106e+05, + "cpu_time": 7.9341768146067532e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3283329367483463e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1780, + "real_time": 3.9483969100800686e+05, + "cpu_time": 7.9405486460677558e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3278502945373039e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1780, + "real_time": 3.9264099884011754e+05, + "cpu_time": 7.8972692415729573e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3352859267085576e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9380591564806539e+05, + "cpu_time": 7.9194736213483172e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3313444206749382e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time_median", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9425793607795110e+05, + "cpu_time": 7.9301382134831371e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3298096297453854e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1056318025670630e+03, + "cpu_time": 2.1532591986926718e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7406532120045559e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8075550890280152e-03, + "cpu_time": 2.7189423207221596e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8096810666831014e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1282, + "real_time": 5.3370554238418338e+05, + "cpu_time": 1.0687280085803487e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9647088454726813e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1282, + "real_time": 5.3548739915938873e+05, + "cpu_time": 1.0688587917317213e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9581711944035673e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1282, + "real_time": 5.3368358152363496e+05, + "cpu_time": 1.0685629219969132e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9647896924360642e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1282, + "real_time": 5.3608136652211857e+05, + "cpu_time": 1.0700690967238075e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9560015801383691e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1282, + "real_time": 5.3363714420546126e+05, + "cpu_time": 1.0688525156006278e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9649606692225997e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3451900675895740e+05, + "cpu_time": 1.0690142669266840e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9617263963346562e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time_median", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3370554238418327e+05, + "cpu_time": 1.0688525156006278e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9647088454726813e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1743182114002141e+03, + "cpu_time": 6.0180104990449968e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3055861457227869e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1969624962836459e-03, + "cpu_time": 5.6294950266157022e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1947944187158124e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 902, + "real_time": 7.7461842012575141e+05, + "cpu_time": 1.5524207882483697e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7073355674391899e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 902, + "real_time": 7.7366675397364656e+05, + "cpu_time": 1.5512527771619172e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7106657863075695e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 902, + "real_time": 7.7464627371760213e+05, + "cpu_time": 1.5534486662971291e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7072382210470920e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 902, + "real_time": 7.7428830679900490e+05, + "cpu_time": 1.5521638747229457e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7084898242488809e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 902, + "real_time": 7.7387926091559790e+05, + "cpu_time": 1.5518018458980366e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7099214385443039e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7421980310632079e+05, + "cpu_time": 1.5522175904656800e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7087301675174074e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time_median", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7428830679900490e+05, + "cpu_time": 1.5521638747229460e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7084898242488809e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3797767996040949e+02, + "cpu_time": 8.1600102285865421e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5325081801543508e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.6570198566758135e-04, + "cpu_time": 5.2570015175117698e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6576627621750817e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 524, + "real_time": 1.3333029941947178e+06, + "cpu_time": 2.6911354179387344e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1457995806371503e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 524, + "real_time": 1.3362325057885004e+06, + "cpu_time": 2.6964610400765310e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1389028345220308e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 524, + "real_time": 1.3372660856711755e+06, + "cpu_time": 2.6985391526717846e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1364767602663560e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 524, + "real_time": 1.3342886889452925e+06, + "cpu_time": 2.6921738034349857e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1434756471745610e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 524, + "real_time": 1.3339588823390570e+06, + "cpu_time": 2.6917124389312956e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1442528368231359e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3350098313877487e+06, + "cpu_time": 2.6940043706106665e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1417815318846469e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time_median", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3342886889452927e+06, + "cpu_time": 2.6921738034349857e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1434756471745610e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6675917974019671e+03, + "cpu_time": 3.2952348034845368e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9226773677986823e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2491232335483985e-03, + "cpu_time": 1.2231735179915784e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2485519212552003e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 280, + "real_time": 2.4996534006537073e+06, + "cpu_time": 5.0262037357141152e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3559084622716970e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 280, + "real_time": 2.5323262172085899e+06, + "cpu_time": 5.0860909964285381e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3126095457191339e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 280, + "real_time": 2.5301398925616271e+06, + "cpu_time": 5.0834442464288743e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3154720119080043e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 280, + "real_time": 2.5321358981143152e+06, + "cpu_time": 5.0852489392858455e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3128585263717508e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 280, + "real_time": 2.5296513373697442e+06, + "cpu_time": 5.0839266892856881e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3161123337741175e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5247813491815967e+06, + "cpu_time": 5.0729829214286115e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3225921760089412e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time_median", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5301398925616271e+06, + "cpu_time": 5.0839266892856872e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3154720119080043e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4096621872562047e+04, + "cpu_time": 2.6171427244421859e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8688611942241188e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.5833040263591309e-03, + "cpu_time": 5.1589819342525347e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6247083458463238e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 143, + "real_time": 4.8887445428571505e+06, + "cpu_time": 9.7742440419580340e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4318045978722420e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 143, + "real_time": 4.8867170926886834e+06, + "cpu_time": 9.7756836713285018e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4332284193618283e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 143, + "real_time": 4.8900380033765221e+06, + "cpu_time": 9.7789685244755987e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4308968536472521e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 143, + "real_time": 4.8905627417762382e+06, + "cpu_time": 9.7836131538464837e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4305287317317119e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 143, + "real_time": 4.8853788830965133e+06, + "cpu_time": 9.7765085874130707e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4341688539346309e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8882882527590226e+06, + "cpu_time": 9.7778035958043393e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4321254913095331e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time_median", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.8887445428571505e+06, + "cpu_time": 9.7765085874130707e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.4318045978722420e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2009148054772932e+03, + "cpu_time": 3.6725186652165075e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5454728535694182e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.5024243491268427e-04, + "cpu_time": 3.7559750809398418e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5029613791299351e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 72, + "real_time": 9.6061401224384699e+06, + "cpu_time": 1.9146057736110024e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4930192119123883e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 72, + "real_time": 9.6047757478016950e+06, + "cpu_time": 1.9133488097221769e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4935154011981816e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 72, + "real_time": 9.6215759404003620e+06, + "cpu_time": 1.9153397486112073e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4874153888976912e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 72, + "real_time": 9.6091668173256852e+06, + "cpu_time": 1.9141139236111827e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4919189808943801e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 72, + "real_time": 9.6088400524523519e+06, + "cpu_time": 1.9142189819444552e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4920377295109925e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6100997360837143e+06, + "cpu_time": 1.9143254475000050e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4915813424827271e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time_median", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.6088400524523519e+06, + "cpu_time": 1.9142189819444552e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4920377295109925e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6742751867628203e+03, + "cpu_time": 7.2754520305726228e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4230925324469144e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.9450633916966046e-04, + "cpu_time": 3.8005303852978233e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9398140692433297e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 34, + "real_time": 2.0629645215676110e+07, + "cpu_time": 4.1014664558825672e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.2530304471259222e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 34, + "real_time": 2.0510392756584812e+07, + "cpu_time": 4.0887099499998920e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.2719443648125587e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 34, + "real_time": 2.0542919800123747e+07, + "cpu_time": 4.0919712411764793e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.2667636661656899e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 34, + "real_time": 2.0581216660930831e+07, + "cpu_time": 4.0949044176470041e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.2606849782302837e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 34, + "real_time": 2.0591362817760777e+07, + "cpu_time": 4.0980867970588140e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.2590783132681360e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0571107450215258e+07, + "cpu_time": 4.0950277723529518e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.2623003539205179e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time_median", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0581216660930831e+07, + "cpu_time": 4.0949044176470049e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.2606849782302837e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5880180036624304e+04, + "cpu_time": 5.0021684355874742e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.2775383692234838e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2303213450058671e-03, + "cpu_time": 1.2215224691170510e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2307996136767707e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18, + "real_time": 3.7948468493090734e+07, + "cpu_time": 7.5104685166669086e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5368417575122161e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18, + "real_time": 3.7882761615845889e+07, + "cpu_time": 7.4784475888887107e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5429763373918962e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18, + "real_time": 3.7798263132572174e+07, + "cpu_time": 7.4877012722222641e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5508967046779346e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18, + "real_time": 3.7883686108721629e+07, + "cpu_time": 7.4935604833336174e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5428898765239277e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18, + "real_time": 3.7976041022274226e+07, + "cpu_time": 7.5170267055556104e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5342738312631583e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7897844074500933e+07, + "cpu_time": 7.4974409133334234e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5415757014738269e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time_median", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7883686108721629e+07, + "cpu_time": 7.4935604833336174e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.5428898765239277e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8959928487404395e+04, + "cpu_time": 1.5999893945024014e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4478188512519486e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8196266877830967e-03, + "cpu_time": 2.1340473542872289e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8206073778314800e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9, + "real_time": 7.5686482919587031e+07, + "cpu_time": 1.4742840155555665e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5466763105533490e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9, + "real_time": 7.5437342955006495e+07, + "cpu_time": 1.4715383355556041e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5583895917450914e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9, + "real_time": 7.5528899828592941e+07, + "cpu_time": 1.4732013311110374e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5540760769611864e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9, + "real_time": 7.5280667179160655e+07, + "cpu_time": 1.4688065666666716e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5657953902181787e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9, + "real_time": 7.5222993062602147e+07, + "cpu_time": 1.4672361566666344e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5685293162504501e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5431277188989863e+07, + "cpu_time": 1.4710132811111030e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5586933371456513e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time_median", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5437342955006495e+07, + "cpu_time": 1.4715383355556041e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5583895917450914e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8759170462199819e+05, + "cpu_time": 2.9536928787552851e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8459867143106200e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.4869220250904028e-03, + "cpu_time": 2.0079308029933404e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4857400951006892e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5, + "real_time": 1.4794081449508667e+08, + "cpu_time": 2.8233237340000415e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6289573896987720e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5, + "real_time": 1.4788330495357513e+08, + "cpu_time": 2.8261807259998477e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6303686353813868e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5, + "real_time": 1.4763534367084503e+08, + "cpu_time": 2.8206401200000072e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6364660294147511e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5, + "real_time": 1.4812240898609161e+08, + "cpu_time": 2.8277416559999436e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6245083757070889e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5, + "real_time": 1.4777485132217407e+08, + "cpu_time": 2.8257249539999521e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6330330039008532e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4787134468555453e+08, + "cpu_time": 2.8247222379999590e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6306666868205705e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time_median", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4788330495357513e+08, + "cpu_time": 2.8257249539999521e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6303686353813868e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8235256617981344e+05, + "cpu_time": 2.7779944418170396e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4768057283737473e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2331839313937569e-03, + "cpu_time": 9.8345756069240817e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2330533520536840e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 2.8911456465721130e+08, + "cpu_time": 5.1522424050000381e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7138973793073416e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 2.8797766566276550e+08, + "cpu_time": 5.1386286350003731e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7285593711888714e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 2.8774751722812653e+08, + "cpu_time": 5.1387350249996191e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7315415762518511e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 2.8883863985538483e+08, + "cpu_time": 5.1473340500001544e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7174452301035590e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 2.8901241719722748e+08, + "cpu_time": 5.1535635850001425e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7152100052063107e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time_mean", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8853816092014313e+08, + "cpu_time": 5.1461007400000656e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7213307124115868e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time_median", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8883863985538483e+08, + "cpu_time": 5.1473340500001544e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.7174452301035590e+09, + "dst_gpu": 4.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time_stddev", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.2982542350366875e+05, + "cpu_time": 7.1592818878905941e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1291100155429067e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time_cv", + "family_index": 4, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/4/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1828149922879059e-03, + "cpu_time": 1.3912051569924190e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1844632051728841e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2697, + "real_time": 2.5976254703345112e+05, + "cpu_time": 5.2154636262512952e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9420617471391424e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2697, + "real_time": 2.5918531072476908e+05, + "cpu_time": 5.2125367259918887e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9508411843887004e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2697, + "real_time": 2.5999985999073097e+05, + "cpu_time": 5.2141775454208732e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9384636593131470e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2697, + "real_time": 2.5964846362231707e+05, + "cpu_time": 5.2097619651462726e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9437937960977252e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2697, + "real_time": 2.5909418691217992e+05, + "cpu_time": 5.2122385539486952e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9522307011352796e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5953807365668967e+05, + "cpu_time": 5.2128356833518052e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9454782176147993e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time_median", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5964846362231710e+05, + "cpu_time": 5.2125367259918881e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9437937960977252e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8642527989137517e+02, + "cpu_time": 2.1564067508940693e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8751812914598249e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4888963089189321e-03, + "cpu_time": 4.1367249648420148e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4890923146476294e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2692, + "real_time": 2.5897404285532492e+05, + "cpu_time": 5.2137402414561744e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9081284649987454e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2692, + "real_time": 2.5913767025523706e+05, + "cpu_time": 5.2064162704308273e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9031350323664909e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2692, + "real_time": 2.5981179458251083e+05, + "cpu_time": 5.2134088930161367e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8826290518908594e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2692, + "real_time": 2.5951443440054252e+05, + "cpu_time": 5.2112523031207465e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8916612277491055e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2692, + "real_time": 2.5897208306046529e+05, + "cpu_time": 5.2123088224368542e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9081883104822123e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5928200503081613e+05, + "cpu_time": 5.2114253060921479e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8987484174974831e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time_median", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5913767025523708e+05, + "cpu_time": 5.2123088224368542e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.9031350323664909e+06, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6960290772187170e+02, + "cpu_time": 2.9662481450342273e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1251682838545476e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4254861523380451e-03, + "cpu_time": 5.6918174411263033e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4244893296791804e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2693, + "real_time": 2.5917411573441140e+05, + "cpu_time": 5.2114975232082012e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5804047361725640e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2693, + "real_time": 2.5971918120877206e+05, + "cpu_time": 5.2113490345337644e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5770879843901407e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2693, + "real_time": 2.5964437518917257e+05, + "cpu_time": 5.2112324136650370e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5775423584723230e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2693, + "real_time": 2.5908630406350864e+05, + "cpu_time": 5.2147729372445890e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5809403800040184e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2693, + "real_time": 2.5974781119687960e+05, + "cpu_time": 5.2115274118083867e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5769141542045094e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5947435747854886e+05, + "cpu_time": 5.2120758640919963e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5785779226487113e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time_median", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5964437518917251e+05, + "cpu_time": 5.2114975232082017e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5775423584723230e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1794345985809144e+02, + "cpu_time": 1.5123798697329639e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9351283559253246e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2253367267105626e-03, + "cpu_time": 2.9016842984814036e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2258681235566464e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2673, + "real_time": 2.6153909517676849e+05, + "cpu_time": 5.2473021436586895e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1322277055608869e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2673, + "real_time": 2.6219454171499069e+05, + "cpu_time": 5.2483576169096457e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1243976119475525e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2673, + "real_time": 2.6213755821148920e+05, + "cpu_time": 5.2482122334457713e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1250767939902760e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2673, + "real_time": 2.6159147426483923e+05, + "cpu_time": 5.2471168237933470e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1316005320977293e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2673, + "real_time": 2.6138161112729629e+05, + "cpu_time": 5.2391220800597116e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1341148922715873e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6176885609907680e+05, + "cpu_time": 5.2460221795734332e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1294835071736068e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time_median", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6159147426483926e+05, + "cpu_time": 5.2473021436586895e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1316005320977293e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.7126895773648857e+02, + "cpu_time": 3.8954630919973772e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4369921839715920e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4183083628404103e-03, + "cpu_time": 7.4255558948363551e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4178033448014116e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2649, + "real_time": 2.6409956265130860e+05, + "cpu_time": 5.3011146055119392e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2037209889786296e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2649, + "real_time": 2.6471420513930428e+05, + "cpu_time": 5.3003290977723768e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1893165088658601e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2649, + "real_time": 2.6487407680877222e+05, + "cpu_time": 5.3018533937335189e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1855807851776108e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2649, + "real_time": 2.6408597298628144e+05, + "cpu_time": 5.2963718195548689e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2040402277826034e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2649, + "real_time": 2.6428934354398353e+05, + "cpu_time": 5.2946897923738603e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1992662209906109e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6441263222593005e+05, + "cpu_time": 5.2988717417893128e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1963849463590637e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time_median", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6428934354398359e+05, + "cpu_time": 5.3003290977723780e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1992662209906109e+07, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6186884422269640e+02, + "cpu_time": 3.1536837399065394e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4766636405486803e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3685762332016494e-03, + "cpu_time": 5.9516136520821120e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3680014579354834e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2605, + "real_time": 2.6867963607608917e+05, + "cpu_time": 5.3827026602687465e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2195937317229436e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2605, + "real_time": 2.6928469956085633e+05, + "cpu_time": 5.3898495547027094e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2168533917239766e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2605, + "real_time": 2.6912659802384704e+05, + "cpu_time": 5.3873104145871825e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2175682463424318e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2605, + "real_time": 2.6773886519149947e+05, + "cpu_time": 5.3811010134355433e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2238790948995313e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2605, + "real_time": 2.6822672599806823e+05, + "cpu_time": 5.3836037159306568e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2216530578028975e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6861130497007206e+05, + "cpu_time": 5.3849134717849689e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2199095044983561e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time_median", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6867963607608911e+05, + "cpu_time": 5.3836037159306568e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2195937317229436e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.3888741165392003e+02, + "cpu_time": 3.5793936624947781e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9033882606361830e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3784829596993438e-03, + "cpu_time": 6.6470773973426468e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3800029837705845e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2519, + "real_time": 2.7654531185188761e+05, + "cpu_time": 5.5584987892025558e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3698105587521160e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2519, + "real_time": 2.7730237166576588e+05, + "cpu_time": 5.5654642794758338e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3633407679250184e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2519, + "real_time": 2.7669794091166585e+05, + "cpu_time": 5.5587715244145750e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3685033500455996e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2519, + "real_time": 2.7751566956071876e+05, + "cpu_time": 5.5653008614524745e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3615243097349182e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2519, + "real_time": 2.7678635313786397e+05, + "cpu_time": 5.5596529297344398e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3677467930421162e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7696952942558040e+05, + "cpu_time": 5.5615376768559753e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3661851558999538e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time_median", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7678635313786392e+05, + "cpu_time": 5.5596529297344398e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3677467930421162e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1723117907238111e+02, + "cpu_time": 3.5361870581397989e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5626643970489275e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5064154527672977e-03, + "cpu_time": 6.3582902132542252e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5056574876085322e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2364, + "real_time": 2.9266794494238199e+05, + "cpu_time": 5.8918929822335695e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4785225804556203e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2364, + "real_time": 2.9607832903757866e+05, + "cpu_time": 5.8968948265648226e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4269366294405204e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2364, + "real_time": 2.9301258427872561e+05, + "cpu_time": 5.8907047377327387e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4732549737631381e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2364, + "real_time": 2.9270706987169973e+05, + "cpu_time": 5.8936309010154800e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4779239550808209e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2364, + "real_time": 2.9547526540268958e+05, + "cpu_time": 5.8893896192894445e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4359719863985240e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9398823870661517e+05, + "cpu_time": 5.8925026133672113e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4585220250277245e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time_median", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9301258427872561e+05, + "cpu_time": 5.8918929822335695e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4732549737631381e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6519834106600351e+03, + "cpu_time": 2.9094376089635136e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4998291652830997e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.6192159860810895e-03, + "cpu_time": 4.9375245118490419e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.6068561537891137e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2149, + "real_time": 3.2484616828898620e+05, + "cpu_time": 6.5287520195441321e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0697888905617082e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2149, + "real_time": 3.2455215529937833e+05, + "cpu_time": 6.5312984690551471e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0770993419590509e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2149, + "real_time": 3.2518999643219489e+05, + "cpu_time": 6.5284608282918856e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0612565846458757e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2149, + "real_time": 3.2514345868951746e+05, + "cpu_time": 6.5343097347608535e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0624103912951159e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2149, + "real_time": 3.2492181840773090e+05, + "cpu_time": 6.5317973941367131e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0679100370861030e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2493071942356165e+05, + "cpu_time": 6.5309236891577474e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0676930491095722e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time_median", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2492181840773090e+05, + "cpu_time": 6.5312984690551483e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0679100370861030e+08, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5642809955884411e+02, + "cpu_time": 2.4057721058940683e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.3689192301363032e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.8917776692138096e-04, + "cpu_time": 3.6836628636282930e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.8943499602271536e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1799, + "real_time": 3.8950698020008119e+05, + "cpu_time": 7.8251872984988661e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3460297931777365e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1799, + "real_time": 3.8978180333543848e+05, + "cpu_time": 7.8262195942192455e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3450807490590014e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1799, + "real_time": 3.8917514733623830e+05, + "cpu_time": 7.8189306225675193e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3471774947309966e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1799, + "real_time": 3.8902902913685178e+05, + "cpu_time": 7.8157796331294777e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3476834907750986e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1799, + "real_time": 3.8936982835860940e+05, + "cpu_time": 7.8179260867159115e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3465039194488666e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8937255767344392e+05, + "cpu_time": 7.8208086470262054e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3464950894383402e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time_median", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8936982835860946e+05, + "cpu_time": 7.8189306225675205e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3465039194488666e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9263525645183893e+02, + "cpu_time": 4.6254202195131791e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0117944464761605e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.5155593450235928e-04, + "cpu_time": 5.9142480378572559e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.5142824835566801e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1311, + "real_time": 5.3160225825659884e+05, + "cpu_time": 1.0599402471395377e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9724822152539902e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1311, + "real_time": 5.2825687386551592e+05, + "cpu_time": 1.0593014012204451e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9849736972224374e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1311, + "real_time": 5.3120500841984292e+05, + "cpu_time": 1.0598609290618219e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9739572921557395e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1311, + "real_time": 5.2853804825812951e+05, + "cpu_time": 1.0597870633103210e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9839177206933880e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1311, + "real_time": 5.3133411918327061e+05, + "cpu_time": 1.0598056971777331e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9734776332673631e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3018726159667154e+05, + "cpu_time": 1.0597390675819716e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9777617117185838e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time_median", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3120500841984292e+05, + "cpu_time": 1.0598056971777331e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9739572921557395e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6431378439039406e+03, + "cpu_time": 2.5180268609596556e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1361470139885014e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.0991650741581220e-03, + "cpu_time": 2.3760819412887069e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1025714461103975e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 905, + "real_time": 7.7030222910554125e+05, + "cpu_time": 1.5439628375689150e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7225054280774560e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 905, + "real_time": 7.6999422915290156e+05, + "cpu_time": 1.5477521613259979e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7235944382429366e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 905, + "real_time": 7.6975912821751856e+05, + "cpu_time": 1.5433939745857141e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7244262823569751e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 905, + "real_time": 7.7024650990252919e+05, + "cpu_time": 1.5439796596685925e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7227023726019664e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 905, + "real_time": 7.7005087504760257e+05, + "cpu_time": 1.5439550276243498e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7233940872677531e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7007059428521863e+05, + "cpu_time": 1.5446087321547139e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7233245217094178e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time_median", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7005087504760246e+05, + "cpu_time": 1.5439628375689150e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7233940872677531e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1669828534094830e+02, + "cpu_time": 1.7746143613061304e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.6641286335760308e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8140054554620170e-04, + "cpu_time": 1.1489086681716219e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8142546260940263e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 526, + "real_time": 1.3292459620538207e+06, + "cpu_time": 2.6809538060837365e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1554009714796290e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 526, + "real_time": 1.3283397991320000e+06, + "cpu_time": 2.6817251387835587e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1575535135970159e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 526, + "real_time": 1.3295845068838913e+06, + "cpu_time": 2.6810334106462523e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1545975289905176e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 526, + "real_time": 1.3303194525244401e+06, + "cpu_time": 2.6826071673003715e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1528547463098483e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 526, + "real_time": 1.3294768167355102e+06, + "cpu_time": 2.6820382281368761e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1548530573846226e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3293933074659326e+06, + "cpu_time": 2.6816715501901591e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1550519635523267e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time_median", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3294768167355100e+06, + "cpu_time": 2.6817251387835587e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1548530573846226e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1289307817086387e+02, + "cpu_time": 6.9553621564728803e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6921434038520495e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.3625445093429046e-04, + "cpu_time": 2.5936666837442007e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3632822007369937e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 278, + "real_time": 2.5317815020211418e+06, + "cpu_time": 5.0965269244603636e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3133222567995329e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 278, + "real_time": 2.5395191133879931e+06, + "cpu_time": 5.0967522482018163e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3032269596934400e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 278, + "real_time": 2.5331179113897812e+06, + "cpu_time": 5.0986236942450153e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3115742312199106e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 278, + "real_time": 2.5407421417708555e+06, + "cpu_time": 5.0990949712237166e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3016368965932441e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 278, + "real_time": 2.5322781992982402e+06, + "cpu_time": 5.0975766258994332e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3126723605347548e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5354877735736025e+06, + "cpu_time": 5.0977148928060690e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3084865409681764e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time_median", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5331179113897812e+06, + "cpu_time": 5.0975766258994332e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3115742312199106e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2870196075852373e+03, + "cpu_time": 1.1277862874689649e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5906146214097068e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6908066575067591e-03, + "cpu_time": 2.2123369219029980e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6897800707914325e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 139, + "real_time": 5.0062860174627183e+06, + "cpu_time": 9.9820455395680293e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3512300219121346e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 139, + "real_time": 5.0095413248691196e+06, + "cpu_time": 9.9843036043154355e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3490523207607884e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 139, + "real_time": 5.0066958094457928e+06, + "cpu_time": 9.9853982374108396e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3509557277970767e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 139, + "real_time": 5.0047827020233916e+06, + "cpu_time": 9.9815952086318098e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3522366501980424e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 139, + "real_time": 5.0097069830345595e+06, + "cpu_time": 9.9849316762586795e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3489415761872439e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0074025673671160e+06, + "cpu_time": 9.9836548532369584e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3504832593710575e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time_median", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0066958094457928e+06, + "cpu_time": 9.9843036043154337e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3509557277970767e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1502616157373263e+03, + "cpu_time": 1.7264537202533370e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4387175231295405e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.2941656613559040e-04, + "cpu_time": 1.7292802542082837e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2940597273708278e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 70, + "real_time": 9.8570393531450201e+06, + "cpu_time": 1.9678458528570835e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.4041085561146727e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 70, + "real_time": 9.9099653240825441e+06, + "cpu_time": 1.9696221857144076e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3859282956781120e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 70, + "real_time": 9.8838508927396368e+06, + "cpu_time": 1.9675551071428083e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3948743626482692e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 70, + "real_time": 9.9017824977636337e+06, + "cpu_time": 1.9713525157141831e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3887264245178514e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 70, + "real_time": 9.9012109584042002e+06, + "cpu_time": 1.9685132357143890e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3889220359979119e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8907698052270077e+06, + "cpu_time": 1.9689777794285741e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3925119349913635e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time_median", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.9012109584042002e+06, + "cpu_time": 1.9685132357143886e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3889220359979119e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1120959780887450e+04, + "cpu_time": 1.5472303435538201e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.2564532806599122e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1354212257296279e-03, + "cpu_time": 7.8580386214558947e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1389617545084288e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1216636128497846e+07, + "cpu_time": 4.2097312757574186e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1630303500308633e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1234958460836701e+07, + "cpu_time": 4.2337535909093112e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1603011667655401e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1192151319348451e+07, + "cpu_time": 4.2153442575757556e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1666848253735123e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1252844877766840e+07, + "cpu_time": 4.2343649515149795e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1576414539309206e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1282687275247142e+07, + "cpu_time": 4.2431074666665435e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1532138367719693e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1235855612339396e+07, + "cpu_time": 4.2272603084848024e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1601743265745616e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time_median", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1234958460836701e+07, + "cpu_time": 4.2337535909093112e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1603011667655401e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4503979683888538e+04, + "cpu_time": 1.4080397271345387e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1337952611143347e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6247981863202867e-03, + "cpu_time": 3.3308564516558695e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6245291337073353e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18, + "real_time": 3.9362040244870715e+07, + "cpu_time": 7.7649587944443330e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4098265020063324e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18, + "real_time": 3.9342004391882151e+07, + "cpu_time": 7.7552584444434717e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4115630373854194e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18, + "real_time": 3.9345426691903010e+07, + "cpu_time": 7.7581707944444954e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4112662966143661e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18, + "real_time": 3.9334075525403023e+07, + "cpu_time": 7.7491898388896316e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4122507318957710e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18, + "real_time": 3.9381702947947711e+07, + "cpu_time": 7.7650627166665092e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4081240259569435e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9353049960401326e+07, + "cpu_time": 7.7585281177776888e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4106061187717667e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time_median", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9345426691903010e+07, + "cpu_time": 7.7581707944444939e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4112662966143661e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8986619624808289e+04, + "cpu_time": 6.7467859253988689e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6450583104558939e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.8246882119463199e-04, + "cpu_time": 8.6959611707012566e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8233605792284072e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9, + "real_time": 7.5541586511664927e+07, + "cpu_time": 1.4762311111111456e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5534791946492805e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9, + "real_time": 7.5589852200614080e+07, + "cpu_time": 1.4754424655555138e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5512102244568124e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9, + "real_time": 7.5746830966737539e+07, + "cpu_time": 1.4771649400000164e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5438506479284554e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9, + "real_time": 7.5536019272274435e+07, + "cpu_time": 1.4752883000000513e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5537410971103354e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9, + "real_time": 7.5571503904130727e+07, + "cpu_time": 1.4745179433331639e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5520724364640751e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5597158571084350e+07, + "cpu_time": 1.4757289519999784e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5508707201217914e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time_median", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5571503904130727e+07, + "cpu_time": 1.4754424655555138e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5520724364640751e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6515372404182694e+04, + "cpu_time": 1.0071017781204016e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0584480533327023e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1444262461641583e-03, + "cpu_time": 6.8244359965665032e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1429444700238203e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5, + "real_time": 1.4837004542350769e+08, + "cpu_time": 2.8344401559998006e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6184589043398538e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5, + "real_time": 1.4824800193309784e+08, + "cpu_time": 2.8321466759998655e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6214377596959586e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5, + "real_time": 1.4852422475814819e+08, + "cpu_time": 2.8313348140000015e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6147026713939924e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5, + "real_time": 1.4847523272037506e+08, + "cpu_time": 2.8337657920001221e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6158954066843896e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5, + "real_time": 1.4823980629444122e+08, + "cpu_time": 2.8289774139998370e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6216379757920117e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4837146222591403e+08, + "cpu_time": 2.8321329703999251e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6184265435812416e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time_median", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4837004542350769e+08, + "cpu_time": 2.8321466759998655e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6184589043398538e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2911471864105125e+05, + "cpu_time": 2.1554575761041295e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1486363957497538e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.7021261841079673e-04, + "cpu_time": 7.6107216667858555e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.7016728343847335e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 2.9170741140842438e+08, + "cpu_time": 5.1786200649996769e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6808863333836803e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 2.9056742787361145e+08, + "cpu_time": 5.1738223600000310e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6953275591063395e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 2.9054909944534302e+08, + "cpu_time": 5.1716404299997973e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6955606678863177e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 2.9100126028060913e+08, + "cpu_time": 5.1787450000006175e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6898184666437640e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 2.9108630120754242e+08, + "cpu_time": 5.1794064850002998e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6887404853670177e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time_mean", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9098230004310608e+08, + "cpu_time": 5.1764468680000859e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6900667024774241e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time_median", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9100126028060913e+08, + "cpu_time": 5.1786200649996775e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6898184666437640e+09, + "dst_gpu": 5.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time_stddev", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7348281379339797e+05, + "cpu_time": 3.4911823629138211e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9991648986838162e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time_cv", + "family_index": 5, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/5/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6271876802240418e-03, + "cpu_time": 6.7443604695253739e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6257605572972754e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2645, + "real_time": 2.6495531601884135e+05, + "cpu_time": 5.3121409792062733e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8648026217642752e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2645, + "real_time": 2.6482284037811798e+05, + "cpu_time": 5.3106315160688770e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8667359603043213e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2645, + "real_time": 2.6456696229559038e+05, + "cpu_time": 5.3078760491495347e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8704757053374061e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2645, + "real_time": 2.6419354394291283e+05, + "cpu_time": 5.2973155879016989e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8759463411462731e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2645, + "real_time": 2.6435632244789432e+05, + "cpu_time": 5.3056562948960811e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8735597110669990e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6457899701667140e+05, + "cpu_time": 5.3067240854444937e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8703040679238550e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time_median", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6456696229559038e+05, + "cpu_time": 5.3078760491495347e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8704757053374061e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1598197989439268e+02, + "cpu_time": 5.8223228858452217e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6222619088286847e+03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1942821745389049e-03, + "cpu_time": 1.0971595266870825e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1942890862598819e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2646, + "real_time": 2.6543235759309249e+05, + "cpu_time": 5.3082115986395907e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7157134065002799e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2646, + "real_time": 2.6567030401039135e+05, + "cpu_time": 5.3108333295539918e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7088028623624248e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2646, + "real_time": 2.6418772979779559e+05, + "cpu_time": 5.3121969160993444e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7520632830582317e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2646, + "real_time": 2.6471084694106004e+05, + "cpu_time": 5.2992994520025968e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7367437854029592e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2646, + "real_time": 2.6360613680061005e+05, + "cpu_time": 5.3025403439156630e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7691666243305011e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6472147502858995e+05, + "cpu_time": 5.3066163280422380e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7364979923308790e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time_median", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6471084694106004e+05, + "cpu_time": 5.3082115986395918e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.7367437854029592e+06, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.5653469683956610e+02, + "cpu_time": 5.5141018007468449e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5042950298211108e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.2356071480300580e-03, + "cpu_time": 1.0390993921320774e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2369878881938520e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2653, + "real_time": 2.6482777551153192e+05, + "cpu_time": 5.3025578326420055e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5466655610758020e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2653, + "real_time": 2.6586503469065053e+05, + "cpu_time": 5.3120061402188428e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5406313224926079e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2653, + "real_time": 2.6431702352695860e+05, + "cpu_time": 5.3032255597439583e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5496542543285090e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2653, + "real_time": 2.6419863679086108e+05, + "cpu_time": 5.2950306445530104e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5503486504521152e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2653, + "real_time": 2.6365608883311827e+05, + "cpu_time": 5.3013437429323583e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5535389370782075e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6457291187062406e+05, + "cpu_time": 5.3028327840180357e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5481677450854484e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time_median", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6431702352695860e+05, + "cpu_time": 5.3025578326420055e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5496542543285090e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.3376585731310729e+02, + "cpu_time": 6.0718810631264955e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8700611820784652e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.1513651621327661e-03, + "cpu_time": 1.1450259343319779e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1456934802692655e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2648, + "real_time": 2.6628535736664874e+05, + "cpu_time": 5.3338231986404327e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0763989732715275e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2648, + "real_time": 2.6687931724121649e+05, + "cpu_time": 5.3396007515103929e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0695522173401445e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2648, + "real_time": 2.6562909925922053e+05, + "cpu_time": 5.3376963557406922e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0839994649854381e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2648, + "real_time": 2.6644904021072778e+05, + "cpu_time": 5.3337608345921664e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0745091044505756e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2648, + "real_time": 2.6507598997390579e+05, + "cpu_time": 5.3206457439579803e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0904345583341684e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6606376081034390e+05, + "cpu_time": 5.3331053768883343e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0789788636763711e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time_median", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6628535736664874e+05, + "cpu_time": 5.3338231986404327e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.0763989732715275e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.1191632782308602e+02, + "cpu_time": 7.4069954079415970e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2452353025095654e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.6757357922582914e-03, + "cpu_time": 1.3888710018820779e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6779122779246904e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2609, + "real_time": 2.6921206982758135e+05, + "cpu_time": 5.3916851054045104e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.0859084106047854e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2609, + "real_time": 2.6931420471747691e+05, + "cpu_time": 5.3918878267537104e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.0836003868372172e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2609, + "real_time": 2.6870025891200959e+05, + "cpu_time": 5.3935110617091565e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.0975006374538757e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2609, + "real_time": 2.6775919832289219e+05, + "cpu_time": 5.3826741433498601e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1189307790810049e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2609, + "real_time": 2.6862867625059694e+05, + "cpu_time": 5.3911707665768254e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.0991254651888974e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6872288160611142e+05, + "cpu_time": 5.3901857807588135e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.0970131358331561e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time_median", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6870025891200959e+05, + "cpu_time": 5.3916851054045104e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.0975006374538757e+07, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1786730849882406e+02, + "cpu_time": 4.2894440773747465e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4038335342242327e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.2992731575589517e-03, + "cpu_time": 7.9578779875948775e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3024938653546118e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2564, + "real_time": 2.7276839166131394e+05, + "cpu_time": 5.4668885335415008e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2013122121820758e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2564, + "real_time": 2.7234243418501585e+05, + "cpu_time": 5.4695573829952814e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2031911258360516e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2564, + "real_time": 2.7290680380349228e+05, + "cpu_time": 5.4750485140403523e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2007029338702285e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2564, + "real_time": 2.7229937759911653e+05, + "cpu_time": 5.4759892160681740e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2033813771047826e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2564, + "real_time": 2.7274767722565442e+05, + "cpu_time": 5.4733748673953593e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2014034485393545e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7261293689491868e+05, + "cpu_time": 5.4721717028081347e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2019982195064986e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time_median", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7274767722565442e+05, + "cpu_time": 5.4733748673953605e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2014034485393545e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7393888417604546e+02, + "cpu_time": 3.8419917013730986e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2081260161920196e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0048638457742667e-03, + "cpu_time": 7.0209633579324958e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0050980081218730e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2499, + "real_time": 2.8097861718588404e+05, + "cpu_time": 5.6522431052416249e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3324194793315551e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2499, + "real_time": 2.8168994364986086e+05, + "cpu_time": 5.6453844817928248e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3265296286707667e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2499, + "real_time": 2.8151708016195905e+05, + "cpu_time": 5.6406270828330971e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3279582170394996e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2499, + "real_time": 2.8236080281564646e+05, + "cpu_time": 5.6488634053622861e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3210020422979352e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2499, + "real_time": 2.8100312846916250e+05, + "cpu_time": 5.6448093997593294e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3322160275234079e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8150991445650265e+05, + "cpu_time": 5.6463854949978332e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3280250789726329e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time_median", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.8151708016195905e+05, + "cpu_time": 5.6453844817928248e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3279582170394996e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.6911810817244077e+02, + "cpu_time": 4.3902463443339821e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7020923674466455e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.0216627512789705e-03, + "cpu_time": 7.7753216605974348e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.0197773683442010e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2372, + "real_time": 2.9527046164337930e+05, + "cpu_time": 5.9782977698147006e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4390488391725975e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2372, + "real_time": 2.9805218391829135e+05, + "cpu_time": 5.9916457293430227e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.3976191778528404e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2372, + "real_time": 2.9491976680529333e+05, + "cpu_time": 5.9882343086003372e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4443273985949546e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2372, + "real_time": 2.9512685026435973e+05, + "cpu_time": 5.9754932166940568e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4412089202521670e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2372, + "real_time": 2.9785130730505032e+05, + "cpu_time": 5.9902266610460135e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4005850162598079e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9624411398727482e+05, + "cpu_time": 5.9847795370996254e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4245578704264736e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time_median", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9527046164337930e+05, + "cpu_time": 5.9882343086003361e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4390488391725975e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5654360261488712e+03, + "cpu_time": 7.3654642605206345e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3337045517631406e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.2842772302848680e-03, + "cpu_time": 1.2306993457089186e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.2744355935798841e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2117, + "real_time": 3.3021911021168478e+05, + "cpu_time": 6.6308269957490405e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9384866560858428e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2117, + "real_time": 3.2970744085506513e+05, + "cpu_time": 6.6253206376949837e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9508063063470531e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2117, + "real_time": 3.3040987219638156e+05, + "cpu_time": 6.6279446291925246e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9339033745393884e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2117, + "real_time": 3.2954630991045863e+05, + "cpu_time": 6.6253922720839782e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9546938356320071e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2117, + "real_time": 3.3019213310548884e+05, + "cpu_time": 6.6268551251770789e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9391352402769375e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3001497325581580e+05, + "cpu_time": 6.6272679319795209e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9434050825762463e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time_median", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.3019213310548884e+05, + "cpu_time": 6.6268551251770789e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 7.9391352402769375e+08, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6851872108758232e+02, + "cpu_time": 2.2696540784187292e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8729784591195756e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1166727298822282e-03, + "cpu_time": 3.4247205661727312e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1170245463853199e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1781, + "real_time": 3.9336144545242080e+05, + "cpu_time": 7.9018883043244004e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3328403331368568e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1781, + "real_time": 3.9343816532289574e+05, + "cpu_time": 7.8994628467145120e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3325804312088418e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1781, + "real_time": 3.9518398035943595e+05, + "cpu_time": 7.9307798652440414e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3266934543326850e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1781, + "real_time": 3.9340897187682631e+05, + "cpu_time": 7.8993312184162636e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3326793171462064e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1781, + "real_time": 3.9371989530410274e+05, + "cpu_time": 7.9023375238637719e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3316268907240479e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9382249166313634e+05, + "cpu_time": 7.9067599517125986e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3312840853097277e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time_median", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9343816532289574e+05, + "cpu_time": 7.9018883043243992e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3325804312088418e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7386149291026800e+02, + "cpu_time": 1.3497044881359664e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6096216290336037e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9650007536192353e-03, + "cpu_time": 1.7070260086037156e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9602289682794987e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1313, + "real_time": 5.3644657485408359e+05, + "cpu_time": 1.0699971035797205e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9546699506567237e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1313, + "real_time": 5.3369925731207163e+05, + "cpu_time": 1.0697890464585989e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9647319827294850e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1313, + "real_time": 5.3730153341076116e+05, + "cpu_time": 1.0712041462299691e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9515596639817796e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1313, + "real_time": 5.3450499368854216e+05, + "cpu_time": 1.0708068728103405e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9617702591774266e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1313, + "real_time": 5.3722890168754710e+05, + "cpu_time": 1.0712797814165696e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9518235089478731e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3583625219060108e+05, + "cpu_time": 1.0706153900990398e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9569110730986576e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time_median", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3644657485408359e+05, + "cpu_time": 1.0708068728103405e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9546699506567237e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6430164134454176e+03, + "cpu_time": 6.8735676051787982e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0067011563286548e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.0662658726960937e-03, + "cpu_time": 6.4202025010521679e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0694808971658506e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 901, + "real_time": 7.7612014238275052e+05, + "cpu_time": 1.5556080632631369e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7020971180590377e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 901, + "real_time": 7.7597418321672815e+05, + "cpu_time": 1.5548399178689506e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7026053770325875e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 901, + "real_time": 7.7616349630136357e+05, + "cpu_time": 1.5548865660376870e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7019461878760290e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 901, + "real_time": 7.7635917488793610e+05, + "cpu_time": 1.5558609667036827e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7012651718873734e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 901, + "real_time": 7.7514593251986010e+05, + "cpu_time": 1.5535204339624222e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7054931362182808e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7595258586172783e+05, + "cpu_time": 1.5549431895671759e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7026813982146621e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time_median", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7612014238275052e+05, + "cpu_time": 1.5548865660376870e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7020971180590377e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.7144169760361382e+02, + "cpu_time": 9.1138310845785350e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6431117101889329e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.0756508347743722e-04, + "cpu_time": 5.8611987535798029e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0795612508168372e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 523, + "real_time": 1.3365129077447362e+06, + "cpu_time": 2.6957878661565674e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1382442890713034e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 523, + "real_time": 1.3370601888603566e+06, + "cpu_time": 2.6968294359464566e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1369597531544304e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 523, + "real_time": 1.3367726383410238e+06, + "cpu_time": 2.6966964608029062e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1376345383649244e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 523, + "real_time": 1.3366028391373260e+06, + "cpu_time": 2.6959424397707465e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1380331368344989e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 523, + "real_time": 1.3378097261275460e+06, + "cpu_time": 2.6964345372849028e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1352022025889478e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3369516600421979e+06, + "cpu_time": 2.6963381479923162e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1372147840028210e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time_median", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3367726383410238e+06, + "cpu_time": 2.6964345372849023e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1376345383649244e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2304021873291754e+02, + "cpu_time": 4.5783127691852394e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2269169133727027e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.9121849679771438e-04, + "cpu_time": 1.6979742591240773e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.9108476717276600e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 277, + "real_time": 2.5364404951057495e+06, + "cpu_time": 5.0805397364622243e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3072362691679316e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 277, + "real_time": 2.5245600030337214e+06, + "cpu_time": 5.0750901985565266e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3228000086825228e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 277, + "real_time": 2.5341386935381144e+06, + "cpu_time": 5.0756339927795213e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3102402885013332e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 277, + "real_time": 2.5265453205131236e+06, + "cpu_time": 5.0793934981956081e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3201890074532018e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 277, + "real_time": 2.5346978266625949e+06, + "cpu_time": 5.0772157725637518e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3095100772012634e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5312764677706608e+06, + "cpu_time": 5.0775746397115272e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3139951302012506e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time_median", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5341386935381144e+06, + "cpu_time": 5.0772157725637518e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3102402885013332e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3399406425922662e+03, + "cpu_time": 2.3541128313739619e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.9961907413677620e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1095841211273315e-03, + "cpu_time": 4.6362938970163643e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1111047139477541e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 139, + "real_time": 5.0082054598886762e+06, + "cpu_time": 9.9779202517985143e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3499456310989542e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 139, + "real_time": 5.0085818716626372e+06, + "cpu_time": 9.9815361007199995e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3496938714172750e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 139, + "real_time": 5.0095890501259472e+06, + "cpu_time": 9.9797001438847147e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3490204150734081e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 139, + "real_time": 5.0001317038864112e+06, + "cpu_time": 9.9681585755389687e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3553548173460536e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 139, + "real_time": 5.0011758303953158e+06, + "cpu_time": 9.9687565179870985e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3546542991018677e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0055367831917983e+06, + "cpu_time": 9.9752143179858588e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3517338068075123e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time_median", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0082054598886762e+06, + "cpu_time": 9.9779202517985143e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3499456310989542e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.5013347467906515e+03, + "cpu_time": 6.3027043982678915e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0150620228446377e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.9927113549655288e-04, + "cpu_time": 6.3183648965854980e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.9955294681245870e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 70, + "real_time": 9.8694043899221085e+06, + "cpu_time": 1.9702002842856407e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3998436657700701e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 70, + "real_time": 9.8877906400178168e+06, + "cpu_time": 1.9798466414285291e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3935216896885614e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 70, + "real_time": 9.8692214781684540e+06, + "cpu_time": 1.9777276414285034e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3999066769577742e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 70, + "real_time": 9.8786432961268090e+06, + "cpu_time": 1.9732947142857224e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3966639946556153e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 70, + "real_time": 9.8709427219416406e+06, + "cpu_time": 1.9741995242855281e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3993138188729911e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8752005052353647e+06, + "cpu_time": 1.9750537611427851e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3978499691890025e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time_median", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.8709427219416406e+06, + "cpu_time": 1.9741995242855277e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3993138188729911e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0286477514000289e+03, + "cpu_time": 3.7914885357771906e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7608091127059110e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.1301111275093791e-04, + "cpu_time": 1.9196887752480213e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.1251648475958458e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1355072713711046e+07, + "cpu_time": 4.2624428363641337e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1425256612173786e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1357691197684318e+07, + "cpu_time": 4.2651599939393818e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1421403830052662e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1364042366092857e+07, + "cpu_time": 4.2682872242425442e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1412062778207808e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1366942165927455e+07, + "cpu_time": 4.2717919424245395e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1407799711750226e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1357996558601204e+07, + "cpu_time": 4.2682017878784493e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1420954589944530e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1360349000403378e+07, + "cpu_time": 4.2671767569698103e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1417495504425807e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time_median", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1357996558601208e+07, + "cpu_time": 4.2682017878784493e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1420954589944530e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.9383165183031088e+03, + "cpu_time": 3.5376934621322827e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.2628848130753112e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.3119081613366205e-04, + "cpu_time": 8.2904779052191298e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3117325860847767e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18, + "real_time": 3.9464132239421211e+07, + "cpu_time": 7.7928735222220287e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4010054290748663e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18, + "real_time": 3.9459981852107577e+07, + "cpu_time": 7.8006909666656703e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4013631456556630e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18, + "real_time": 3.9384274640017085e+07, + "cpu_time": 7.7850963944443822e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4079014842036905e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18, + "real_time": 3.9422559448414378e+07, + "cpu_time": 7.7915300111107573e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4045919361381898e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18, + "real_time": 3.9437742075986333e+07, + "cpu_time": 7.7961678111111924e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4032812462081919e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9433738051189318e+07, + "cpu_time": 7.7932717411108062e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4036286482561207e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time_median", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9437742075986333e+07, + "cpu_time": 7.7928735222220287e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4032812462081919e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2404344559644138e+04, + "cpu_time": 5.7762579604403385e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7981922580408943e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.2174163954682009e-04, + "cpu_time": 7.4118523674333284e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.2212031546819093e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9, + "real_time": 7.5570177700784475e+07, + "cpu_time": 1.4756075755556393e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5521347728313394e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9, + "real_time": 7.5579351021183863e+07, + "cpu_time": 1.4757624433332241e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5517036382696805e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9, + "real_time": 7.5353554553455770e+07, + "cpu_time": 1.4721090277776563e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5623462966112900e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9, + "real_time": 7.5579990943272904e+07, + "cpu_time": 1.4756561999999776e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5516735666385050e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9, + "real_time": 7.5614248712857559e+07, + "cpu_time": 1.4759508522221649e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5500644464428148e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5539464586310938e+07, + "cpu_time": 1.4750172197777325e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5535845441587257e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time_median", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5579351021183863e+07, + "cpu_time": 1.4756561999999779e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5517036382696805e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0527560803794271e+05, + "cpu_time": 1.6310597833179616e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9611206628180295e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3936504397334647e-03, + "cpu_time": 1.1057903334604749e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3960891041618719e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5, + "real_time": 1.4837051331996918e+08, + "cpu_time": 2.8326186620001864e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6184474932846551e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5, + "real_time": 1.4821850955486298e+08, + "cpu_time": 2.8307316199998242e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6221583499412913e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5, + "real_time": 1.4819179475307465e+08, + "cpu_time": 2.8323438060001534e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6228113229518809e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5, + "real_time": 1.4813576042652130e+08, + "cpu_time": 2.8333854580000663e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6241816996396365e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5, + "real_time": 1.4781432151794434e+08, + "cpu_time": 2.8282334780001295e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6320628913810968e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4814617991447449e+08, + "cpu_time": 2.8314626048000717e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6239323514397125e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time_median", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4819179475307465e+08, + "cpu_time": 2.8323438060001534e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6228113229518809e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0486356939489886e+05, + "cpu_time": 2.0476783742799706e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0163559835402435e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3828474653424583e-03, + "cpu_time": 7.2318750415725731e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3842300288931580e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 2.9652483761310577e+08, + "cpu_time": 5.2629026900001460e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6210856150977044e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 2.9672548174858093e+08, + "cpu_time": 5.2596246500002050e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6186370569609332e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 2.9704859852790833e+08, + "cpu_time": 5.2689574949999952e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6147008581126828e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 2.9597789049148560e+08, + "cpu_time": 5.2544479899995619e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6277771363833961e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 2.9627436399459839e+08, + "cpu_time": 5.2625300950001019e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6241469208573718e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time_mean", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9651023447513580e+08, + "cpu_time": 5.2616925840000027e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6212695174824176e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time_median", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9652483761310577e+08, + "cpu_time": 5.2625300950001019e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6210856150977044e+09, + "dst_gpu": 6.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time_stddev", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1091528757182538e+05, + "cpu_time": 5.2845341866465495e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0184536549339574e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time_cv", + "family_index": 6, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/6/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3858384628753283e-03, + "cpu_time": 1.0043411131079767e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3858271610843512e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2691, + "real_time": 2.6070693310650237e+05, + "cpu_time": 5.2492139130433253e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9277820033335355e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2691, + "real_time": 2.6332634110598976e+05, + "cpu_time": 5.2628542028986081e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.8887108509506700e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2691, + "real_time": 2.5986424133201299e+05, + "cpu_time": 5.2390166926789947e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9405190754648559e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2691, + "real_time": 2.6125507864015075e+05, + "cpu_time": 5.2540437532514718e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9195410299006812e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2691, + "real_time": 2.5879846348412352e+05, + "cpu_time": 5.2484246339655295e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9567468300012499e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6079021153375591e+05, + "cpu_time": 5.2507106391675852e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9266599579301984e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time_median", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6070693310650234e+05, + "cpu_time": 5.2492139130433253e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 3.9277820033335355e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6940684046049241e+03, + "cpu_time": 8.7004336519040021e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5442915775259971e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.4959048679081614e-03, + "cpu_time": 1.6570011660904083e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4795312168236018e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2699, + "real_time": 2.5943509031210691e+05, + "cpu_time": 5.2470466320860095e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8940747665869128e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2699, + "real_time": 2.6115313166172410e+05, + "cpu_time": 5.2434999147834512e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8421422211884782e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2699, + "real_time": 2.5972255176777265e+05, + "cpu_time": 5.2442263542052510e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8853375883631054e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2699, + "real_time": 2.6093449594500251e+05, + "cpu_time": 5.2529039422009967e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8487131131625446e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2699, + "real_time": 2.6062505073469615e+05, + "cpu_time": 5.2459599814747577e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8580320434537437e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6037406408426049e+05, + "cpu_time": 5.2467273649500939e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8656599465509569e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time_median", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6062505073469612e+05, + "cpu_time": 5.2459599814747577e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 7.8580320434537437e+06, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5666447365032627e+02, + "cpu_time": 3.7253323582118270e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.2875068626946239e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.9060669936981881e-03, + "cpu_time": 7.1002971930622921e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.9082198801356541e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2726, + "real_time": 2.6068190593232814e+05, + "cpu_time": 5.2466955429195950e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5712636384756612e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2726, + "real_time": 2.6167407224249633e+05, + "cpu_time": 5.2535411702129920e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5653060178633938e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2726, + "real_time": 2.6088568767553545e+05, + "cpu_time": 5.2440509354359237e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5700363007625818e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2726, + "real_time": 2.6038361841449564e+05, + "cpu_time": 5.2400854732208961e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5730636300935492e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2726, + "real_time": 2.6156906825927601e+05, + "cpu_time": 5.2447267754955136e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5659343924947226e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6103887050482631e+05, + "cpu_time": 5.2458199794569844e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5691207959379820e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time_median", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6088568767553545e+05, + "cpu_time": 5.2447267754955136e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 1.5700363007625818e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.6232439334626247e+02, + "cpu_time": 4.9394476283545964e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3794068659492601e+04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1541787713790527e-03, + "cpu_time": 9.4159686144355619e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1536945241549320e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2672, + "real_time": 2.6224052090763149e+05, + "cpu_time": 5.2749163922157325e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1238498046171337e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2672, + "real_time": 2.6149513406832743e+05, + "cpu_time": 5.2816703181146027e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1327542782725237e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2672, + "real_time": 2.6343100828902511e+05, + "cpu_time": 5.2911698952091078e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1097326215340953e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2672, + "real_time": 2.6346477776194899e+05, + "cpu_time": 5.2741875598801975e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1093340330303282e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2672, + "real_time": 2.6388154852437205e+05, + "cpu_time": 5.2817221744007862e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1044231951077051e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6290259791026101e+05, + "cpu_time": 5.2807332679640851e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1160187865123570e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time_median", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6343100828902511e+05, + "cpu_time": 5.2816703181146015e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 3.1097326215340953e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.9660780181954453e+02, + "cpu_time": 6.8458294778311347e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1832804700215309e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7907871954910299e-03, + "cpu_time": 1.2963785766953631e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7974112195450927e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2645, + "real_time": 2.6683606093376171e+05, + "cpu_time": 5.3348514366737986e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1400996337099642e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2645, + "real_time": 2.6574777563574369e+05, + "cpu_time": 5.3336248393192957e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1652444543721385e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2645, + "real_time": 2.6524508843332221e+05, + "cpu_time": 5.3246372703218344e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1769287027226672e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2645, + "real_time": 2.6579792083140695e+05, + "cpu_time": 5.3338588922498992e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1640813249221057e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2645, + "real_time": 2.6390850584688183e+05, + "cpu_time": 5.3258606729678833e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.2082121784683593e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6550707033622329e+05, + "cpu_time": 5.3305666223065439e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1709132588390470e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time_median", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6574777563574369e+05, + "cpu_time": 5.3336248393192957e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 6.1652444543721385e+07, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0645177316685217e+03, + "cpu_time": 4.8952666162237273e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4777434942781102e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.0093762110382817e-03, + "cpu_time": 9.1833888647761407e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0151974113864890e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2583, + "real_time": 2.6948588049440185e+05, + "cpu_time": 5.4226219279908750e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2159449667597966e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2583, + "real_time": 2.6898017401523609e+05, + "cpu_time": 5.4131371196282981e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2182310506700724e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2583, + "real_time": 2.7090575353307719e+05, + "cpu_time": 5.4223124932246527e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2095719479061960e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2583, + "real_time": 2.6813848127799912e+05, + "cpu_time": 5.4088292140924453e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2220551054000703e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2583, + "real_time": 2.6929230617443653e+05, + "cpu_time": 5.4215502400311956e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2168190196556984e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6936051909903018e+05, + "cpu_time": 5.4176901989934931e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2165244180783668e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time_median", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6929230617443647e+05, + "cpu_time": 5.4215502400311944e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.2168190196556984e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.0057530459122012e+03, + "cpu_time": 6.3212945033662163e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.5348785118236631e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.7338547210864150e-03, + "cpu_time": 1.1667877400115267e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7277332410532286e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2500, + "real_time": 2.7761368296341971e+05, + "cpu_time": 5.5812431760004989e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3606905574836338e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2500, + "real_time": 2.7835697757545859e+05, + "cpu_time": 5.6016861359994439e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3543868226631442e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2500, + "real_time": 2.7789303646422923e+05, + "cpu_time": 5.6029011119999271e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3583174603382292e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2500, + "real_time": 2.7866839324124157e+05, + "cpu_time": 5.5872731879999267e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3517557638216212e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2500, + "real_time": 2.7820197464898229e+05, + "cpu_time": 5.6068593199997849e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3556985920998293e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7814681297866633e+05, + "cpu_time": 5.5959925863999152e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3561698392812917e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time_median", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7820197464898229e+05, + "cpu_time": 5.6016861359994439e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 2.3556985920998293e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0862209629311690e+02, + "cpu_time": 1.1088317588065670e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4617356051553099e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4690878242220159e-03, + "cpu_time": 1.9814746743971553e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4692215932155602e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2355, + "real_time": 2.9712956363627897e+05, + "cpu_time": 5.9242307473463879e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4112742736177987e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2355, + "real_time": 2.9465393909705983e+05, + "cpu_time": 5.9190623397025501e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4483369338844812e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2355, + "real_time": 2.9386840691501292e+05, + "cpu_time": 5.9191994522302062e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4602276704724562e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2355, + "real_time": 2.9693063555657072e+05, + "cpu_time": 5.9187920636935090e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4142295979098588e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2355, + "real_time": 2.9398363978095795e+05, + "cpu_time": 5.9112821698512090e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4584793935356212e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9531323699717608e+05, + "cpu_time": 5.9185133545647725e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4385095738840431e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time_median", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9465393909705989e+05, + "cpu_time": 5.9190623397025501e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 4.4483369338844812e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5972677456013639e+03, + "cpu_time": 4.6321810762221895e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3970425709778289e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.4087238413110398e-03, + "cpu_time": 7.8265956308935694e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.4005573967484521e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2153, + "real_time": 3.2562459170164284e+05, + "cpu_time": 6.5608016720857262e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0504976184413111e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2153, + "real_time": 3.2630174687507277e+05, + "cpu_time": 6.5686823827218555e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0337908855990255e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2153, + "real_time": 3.2657374278534093e+05, + "cpu_time": 6.5636507338599395e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0270997222305453e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2153, + "real_time": 3.2615141305976012e+05, + "cpu_time": 6.5591990339070361e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0374939216335034e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2153, + "real_time": 3.2658689705433504e+05, + "cpu_time": 6.5617480910357216e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0267764066598940e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2624767829523038e+05, + "cpu_time": 6.5628163827220572e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0351317109128571e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time_median", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2630174687507277e+05, + "cpu_time": 6.5617480910357239e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 8.0337908855990255e+08, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9427614947397439e+02, + "cpu_time": 3.6534783534625933e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.7187051277420693e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2085178706380958e-03, + "cpu_time": 5.5669367241191058e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2095265488358676e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1800, + "real_time": 3.8944718421488587e+05, + "cpu_time": 7.8253891166670853e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3462364635064683e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1800, + "real_time": 3.8935412377921038e+05, + "cpu_time": 7.8227893333329726e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3465582306181149e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1800, + "real_time": 3.9092843773283268e+05, + "cpu_time": 7.8646581777775288e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3411354851557448e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1800, + "real_time": 3.8929767843607505e+05, + "cpu_time": 7.8243645555554994e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3467534718065143e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1800, + "real_time": 3.8953882435129746e+05, + "cpu_time": 7.8262505111107812e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3459197574801474e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8971324970286025e+05, + "cpu_time": 7.8326903388887749e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3453206817133980e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time_median", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.8944718421488593e+05, + "cpu_time": 7.8253891166670865e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 1.3462364635064683e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.8548667060273215e+02, + "cpu_time": 1.7916964936085842e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3610290677753207e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7589514113912897e-03, + "cpu_time": 2.2874598842659889e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7549935118579447e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1324, + "real_time": 5.2997230652725359e+05, + "cpu_time": 1.0627290800603435e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9785486658180268e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1324, + "real_time": 5.3278863184003497e+05, + "cpu_time": 1.0630647243201262e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9680900404699802e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1324, + "real_time": 5.2984216527625069e+05, + "cpu_time": 1.0626547522658086e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9790346422378266e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1324, + "real_time": 5.3216095871694409e+05, + "cpu_time": 1.0622367484895394e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9704113629984205e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1324, + "real_time": 5.2983877476792946e+05, + "cpu_time": 1.0628644342900866e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9790473063419693e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.3092056742568256e+05, + "cpu_time": 1.0627099478851811e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9750264035732446e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time_median", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2997230652725371e+05, + "cpu_time": 1.0627290800603435e+06, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 1.9785486658180268e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4370682389338563e+03, + "cpu_time": 3.0702712895285777e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.3397495306053450e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7067481033968698e-03, + "cpu_time": 2.8890962163650519e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7036345037942770e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 899, + "real_time": 7.7039247404087801e+05, + "cpu_time": 1.5438317274749940e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7221865096889853e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 899, + "real_time": 7.7053539430851059e+05, + "cpu_time": 1.5446796785317778e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7216815937209663e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 899, + "real_time": 7.7028551665461052e+05, + "cpu_time": 1.5439076440488801e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7225644967440624e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 899, + "real_time": 7.7136956337628025e+05, + "cpu_time": 1.5455682736371330e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7187383318843660e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 899, + "real_time": 7.7026183526596264e+05, + "cpu_time": 1.5442471768631057e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7226482008885684e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7056895672924840e+05, + "cpu_time": 1.5444469001111784e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7215638265853901e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time_median", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7039247404087801e+05, + "cpu_time": 1.5442471768631057e+06, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 2.7221865096889853e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.6040890729516150e+02, + "cpu_time": 7.1077215487587637e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6249793874384991e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.9749215599004903e-04, + "cpu_time": 4.6021145487404639e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9707561203048453e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 524, + "real_time": 1.3310991675070424e+06, + "cpu_time": 2.6818104751907783e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1510079056358585e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 524, + "real_time": 1.3303935400277858e+06, + "cpu_time": 2.6814272958013690e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1526791688363128e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 524, + "real_time": 1.3308570769576372e+06, + "cpu_time": 2.6809227041982352e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1515810920795894e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 524, + "real_time": 1.3311339947945999e+06, + "cpu_time": 2.6831412671755888e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1509254638539977e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 524, + "real_time": 1.3303929139621812e+06, + "cpu_time": 2.6827879522900605e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1526806524461317e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3307753386498494e+06, + "cpu_time": 2.6820179389312067e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1517748565703783e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time_median", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3308570769576370e+06, + "cpu_time": 2.6818104751907783e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 3.1515810920795894e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.6475996868684945e+02, + "cpu_time": 9.2820075261232159e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.6392610202493588e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.7409582826874452e-04, + "cpu_time": 3.4608297697748171e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7410780951689611e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 280, + "real_time": 2.5043375099942619e+06, + "cpu_time": 5.0236052928572204e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3496315758251052e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 280, + "real_time": 2.5299028917548382e+06, + "cpu_time": 5.0757836607138934e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3157826046759205e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 280, + "real_time": 2.5328125987601066e+06, + "cpu_time": 5.0777330392861590e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3119734180517316e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 280, + "real_time": 2.5301949079481089e+06, + "cpu_time": 5.0752703071428388e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3153999218197937e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 280, + "real_time": 2.5347812009775746e+06, + "cpu_time": 5.0804515642856322e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3094012204149270e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5264058218869781e+06, + "cpu_time": 5.0665687728571482e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3204377481574955e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time_median", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5301949079481084e+06, + "cpu_time": 5.0757836607138934e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 3.3153999218197937e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2497908079423572e+04, + "cpu_time": 2.4103237404757190e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6528607455555353e+07, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.9469123175503362e-03, + "cpu_time": 4.7573098255142901e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.9778398841318575e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 139, + "real_time": 5.0064998397456231e+06, + "cpu_time": 9.9785172733816244e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3510868944424934e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 139, + "real_time": 5.0385806227116277e+06, + "cpu_time": 1.0050597812948627e+07, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3297504309796987e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 139, + "real_time": 5.0084428901938228e+06, + "cpu_time": 9.9789644532376770e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3497868235352359e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 139, + "real_time": 5.0057228688123841e+06, + "cpu_time": 9.9777828848921750e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3516070385215755e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 139, + "real_time": 5.0066564493417311e+06, + "cpu_time": 9.9770427841723952e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3509820715191765e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0131805341610378e+06, + "cpu_time": 9.9925810417264998e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3466426517996364e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time_median", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0066564493417311e+06, + "cpu_time": 9.9785172733816244e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 3.3509820715191765e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4233955017668171e+04, + "cpu_time": 3.2440598402029573e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.4665190516641326e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.8393062888269279e-03, + "cpu_time": 3.2464683815488526e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8286614486830765e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 69, + "real_time": 9.9131423116162196e+06, + "cpu_time": 1.9754072231884547e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3848431652878537e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 69, + "real_time": 9.8967294519146290e+06, + "cpu_time": 1.9714820942028936e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3904566314590459e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 69, + "real_time": 9.8791920269529019e+06, + "cpu_time": 1.9720379333332106e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3964753300123262e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 69, + "real_time": 9.9076000784618258e+06, + "cpu_time": 1.9724568072463270e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3867366197939420e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 69, + "real_time": 9.9150645344153699e+06, + "cpu_time": 1.9754340985509373e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3841869494174199e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.9023456806721911e+06, + "cpu_time": 1.9733636313043647e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3885397391941175e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time_median", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.9076000784618277e+06, + "cpu_time": 1.9724568072463267e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 3.3867366197939420e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4778183941188234e+04, + "cpu_time": 1.9093905607024033e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.0623797430694588e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4923922490438712e-03, + "cpu_time": 9.6758171196269769e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4939708938675228e-03, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 33, + "real_time": 2.1270128531437933e+07, + "cpu_time": 4.2347956666665629e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1550756217016249e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 33, + "real_time": 2.1301716566085815e+07, + "cpu_time": 4.2558018454545878e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1503970016596289e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 33, + "real_time": 2.1298880096186291e+07, + "cpu_time": 4.2528634515147962e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1508165545293765e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 33, + "real_time": 2.1281852415113740e+07, + "cpu_time": 4.2472590666668773e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1533375333597035e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 33, + "real_time": 2.1285207377690256e+07, + "cpu_time": 4.2438085333335981e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1528405060473623e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1287556997302808e+07, + "cpu_time": 4.2469057127272844e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1524934434595394e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time_median", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1285207377690259e+07, + "cpu_time": 4.2472590666668773e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 3.1528405060473623e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2947203839234737e+04, + "cpu_time": 8.2311363697613153e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9175250568584495e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.0820524595072995e-04, + "cpu_time": 1.9381490728870999e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.0825663597706382e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 18, + "real_time": 3.9346071374085218e+07, + "cpu_time": 7.7573423444442198e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4112104032932954e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 18, + "real_time": 3.9341955135265984e+07, + "cpu_time": 7.7515137499997556e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4115673087046890e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 18, + "real_time": 3.9374667116337352e+07, + "cpu_time": 7.7623858111110166e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4087330212452850e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 18, + "real_time": 3.9349325208200350e+07, + "cpu_time": 7.7622549499993622e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4109283269749489e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 18, + "real_time": 3.9361022205816373e+07, + "cpu_time": 7.7626025444437295e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4099146942420278e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9354608207941063e+07, + "cpu_time": 7.7592198799996182e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4104707508920493e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time_median", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9349325208200350e+07, + "cpu_time": 7.7622549499993622e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 3.4109283269749489e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.3270145447655255e+04, + "cpu_time": 4.8370127073688971e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1497612680204529e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 3.3719419534146384e-04, + "cpu_time": 6.2338904969517822e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3712685198068894e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9, + "real_time": 7.5637893544303045e+07, + "cpu_time": 1.4756728088890564e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5489546763061361e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9, + "real_time": 7.5583902498086289e+07, + "cpu_time": 1.4758394877779007e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5514897634029484e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9, + "real_time": 7.5506888329982758e+07, + "cpu_time": 1.4752505055556262e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5551121485350356e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9, + "real_time": 7.5531440476576492e+07, + "cpu_time": 1.4741848311109605e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5539565286491013e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9, + "real_time": 7.5557786557409495e+07, + "cpu_time": 1.4743390744443887e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5527173072498660e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5563582281271622e+07, + "cpu_time": 1.4750573415555868e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5524460848286171e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time_median", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.5557786557409495e+07, + "cpu_time": 1.4752505055556262e+08, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 3.5527173072498660e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0537028896814387e+04, + "cpu_time": 7.5911909492567385e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.3752569321570247e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.6880139044625411e-04, + "cpu_time": 5.1463700667060935e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.6862575122561378e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5, + "real_time": 1.4803273379802704e+08, + "cpu_time": 2.8277869799999279e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6267040283975039e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5, + "real_time": 1.4800633192062378e+08, + "cpu_time": 2.8288510140000653e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6273509723078971e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5, + "real_time": 1.4805721342563629e+08, + "cpu_time": 2.8283978460003710e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6261043928781667e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5, + "real_time": 1.4802939891815186e+08, + "cpu_time": 2.8253278500001216e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6267857325884686e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5, + "real_time": 1.4802623689174652e+08, + "cpu_time": 2.8251264720001924e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6268632052885365e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4803038299083710e+08, + "cpu_time": 2.8270980324001360e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6267616662921143e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time_median", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4802939891815186e+08, + "cpu_time": 2.8277869799999279e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 3.6267857325884686e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8179734830143512e+04, + "cpu_time": 1.7505502580968695e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4539348320333561e+05, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2281083425467334e-04, + "cpu_time": 6.1920394624967998e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2280748617779776e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2, + "real_time": 2.9301244020462036e+08, + "cpu_time": 5.2058296850009358e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6644922763353333e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2, + "real_time": 2.9362787306308746e+08, + "cpu_time": 5.2160496449994296e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6568116398449035e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2, + "real_time": 2.9363411664962769e+08, + "cpu_time": 5.2171576750004077e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6567338845070863e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2, + "real_time": 2.9321302473545074e+08, + "cpu_time": 5.2158673550002277e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6619854284057660e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2, + "real_time": 2.9354733228683472e+08, + "cpu_time": 5.2186278300007415e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6578149616798820e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time_mean", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9340695738792419e+08, + "cpu_time": 5.2147064380003482e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6595676381545944e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time_median", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.9354733228683472e+08, + "cpu_time": 5.2160496449994296e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 3.6578149616798820e+09, + "dst_gpu": 7.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time_stddev", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7989456028732890e+05, + "cpu_time": 5.0826802769406058e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.4926827853900502e+06, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time_cv", + "family_index": 7, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToGPU/0/7/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 9.5394656888544733e-04, + "cpu_time": 9.7468195714764535e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.5439765861283577e-04, + "dst_gpu": 0.0000000000000000e+00, + "src_gpu": NaN + } + ] +} diff --git a/results/prefetch_managed_GPUToHost.json b/results/prefetch_managed_GPUToHost.json new file mode 100644 index 0000000..f3b6454 --- /dev/null +++ b/results/prefetch_managed_GPUToHost.json @@ -0,0 +1,3526 @@ +{ + "context": { + "date": "2023-10-09T17:44:34-04:00", + "host_name": "frontier08051", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1795, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [0.4,3.8,18.46], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3482, + "real_time": 1.9999642918922947e+05, + "cpu_time": 2.7772255600229761e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 5.1200914143878426e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3482, + "real_time": 2.0064990229217662e+05, + "cpu_time": 2.7779026450315927e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 5.1034163899511956e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3482, + "real_time": 2.0005319797832976e+05, + "cpu_time": 2.7767235956346936e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 5.1186384939016178e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3482, + "real_time": 1.9994312684019058e+05, + "cpu_time": 2.7765673779437103e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 5.1214563670320949e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3482, + "real_time": 2.0059431243388570e+05, + "cpu_time": 2.7788827685238374e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 5.1048306782750990e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0024739374676242e+05, + "cpu_time": 2.7774603894313623e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 5.1136866687095705e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0005319797832973e+05, + "cpu_time": 2.7772255600229756e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 5.1186384939016178e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4483302505006520e+02, + "cpu_time": 9.5055664863842637e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.8008059900340231e+03, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7220350217698672e-03, + "cpu_time": 3.4223949772800788e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7210295741985479e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3496, + "real_time": 2.0007919733007241e+05, + "cpu_time": 2.7754034010297502e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.0235946701752288e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3496, + "real_time": 2.0036129535999559e+05, + "cpu_time": 2.7760693278031994e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.0221535034100736e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3496, + "real_time": 1.9988585186402156e+05, + "cpu_time": 2.7763548798627011e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.0245847722094981e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3496, + "real_time": 2.0052423996873360e+05, + "cpu_time": 2.7833651430205978e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.0213229085517697e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3496, + "real_time": 1.9997787785541732e+05, + "cpu_time": 2.7768625457665900e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.0241132779100148e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0016569247564810e+05, + "cpu_time": 2.7776110594965681e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.0231538264513172e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0007919733007238e+05, + "cpu_time": 2.7763548798627011e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 1.0235946701752288e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6827378871437810e+02, + "cpu_time": 3.2595409872632530e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3707117304695670e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3402585897531662e-03, + "cpu_time": 1.1735051875311994e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3396927177839050e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3497, + "real_time": 1.9995599617700346e+05, + "cpu_time": 2.7719630997998297e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.0484506983097278e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3497, + "real_time": 2.0034806178347414e+05, + "cpu_time": 2.7781095967972541e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.0444420392879795e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3497, + "real_time": 2.0046652010764499e+05, + "cpu_time": 2.7769111181012314e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.0432339513852790e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3497, + "real_time": 1.9994550705913344e+05, + "cpu_time": 2.7765685130111565e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.0485581597933166e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3497, + "real_time": 1.9994785090786434e+05, + "cpu_time": 2.7736841778667487e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.0485341459796086e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0013278720702408e+05, + "cpu_time": 2.7754473011152440e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.0466437989511821e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9995599617700343e+05, + "cpu_time": 2.7765685130111565e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 2.0484506983097278e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5409208421493190e+02, + "cpu_time": 2.5365309465203117e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5970117758805984e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2696174762813375e-03, + "cpu_time": 9.1391789190198996e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2689124395810626e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3445, + "real_time": 2.0119223329887411e+05, + "cpu_time": 2.7982265602322167e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.0717277529449463e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3445, + "real_time": 2.0111438774176655e+05, + "cpu_time": 2.7967275036284514e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.0733038008790463e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3445, + "real_time": 2.0147438260087275e+05, + "cpu_time": 2.7959320725689456e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.0660256129081264e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3445, + "real_time": 2.0108099826456135e+05, + "cpu_time": 2.7977571001451381e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.0739801725182526e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3445, + "real_time": 2.0111857121675147e+05, + "cpu_time": 2.7979124499274354e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.0732190719330624e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0119611462456529e+05, + "cpu_time": 2.7973111373004375e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.0716512822366871e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0111857121675144e+05, + "cpu_time": 2.7977571001451381e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 4.0732190719330624e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6077326510130200e+02, + "cpu_time": 9.5405800079309216e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.2506351499553130e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.9908732532587494e-04, + "cpu_time": 3.4106252539136984e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.9835794488020010e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3458, + "real_time": 2.0264279566329657e+05, + "cpu_time": 2.8257444505494426e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.0851628336311653e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3458, + "real_time": 2.0202823927616049e+05, + "cpu_time": 2.8196946096009266e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.1097573580315441e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3458, + "real_time": 2.0253480355642593e+05, + "cpu_time": 2.8201929959514155e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.0894738643945903e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3458, + "real_time": 2.0211948440523315e+05, + "cpu_time": 2.8230495026026561e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.1060962767703354e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3458, + "real_time": 2.0211693962117276e+05, + "cpu_time": 2.8220993290919648e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.1061983378080472e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0228845250445776e+05, + "cpu_time": 2.8221561775592814e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.0993377341271386e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0211948440523312e+05, + "cpu_time": 2.8220993290919642e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 8.1060962767703354e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.7925155634424050e+02, + "cpu_time": 2.4284955871685838e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1175174202780913e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3804621711567384e-03, + "cpu_time": 8.6051069975469916e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3797639473278806e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3424, + "real_time": 2.0507595386376788e+05, + "cpu_time": 2.8857862178738369e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.5978470114427853e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3424, + "real_time": 2.0463651628369489e+05, + "cpu_time": 2.8704693370327051e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.6012782368994471e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3424, + "real_time": 2.0403832853266509e+05, + "cpu_time": 2.8689911623831792e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.6059727716674602e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3424, + "real_time": 2.0413086382585423e+05, + "cpu_time": 2.8712451255841210e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.6052447623968640e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3424, + "real_time": 2.0472080943414834e+05, + "cpu_time": 2.8713428037383256e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.6006189156134778e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0452049438802610e+05, + "cpu_time": 2.8735669293224334e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.6021923396040067e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0463651628369495e+05, + "cpu_time": 2.8712451255841210e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 1.6012782368994471e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.3197752088858653e+02, + "cpu_time": 6.8953107434301455e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.3838615195512952e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1121478421082729e-03, + "cpu_time": 2.3995650399053037e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1120195346756186e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3368, + "real_time": 2.0774718821770456e+05, + "cpu_time": 2.9650384174584190e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.1546034659839940e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3368, + "real_time": 2.0834438722639985e+05, + "cpu_time": 2.9659804869358585e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.1455611006590039e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3368, + "real_time": 2.0771126948511880e+05, + "cpu_time": 2.9606787381235161e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.1551489797569811e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3368, + "real_time": 2.0767317411784004e+05, + "cpu_time": 2.9632445011876535e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.1557277572505778e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3368, + "real_time": 2.0819858989835664e+05, + "cpu_time": 2.9632874881235335e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.1477638744813269e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0793492178908398e+05, + "cpu_time": 2.9636459263657959e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.1517610356263769e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0774718821770453e+05, + "cpu_time": 2.9632874881235335e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 3.1546034659839940e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1263353406789361e+02, + "cpu_time": 2.0299649449925062e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.7357403600704292e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5035162510365107e-03, + "cpu_time": 6.8495528663971446e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5025696131589032e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3290, + "real_time": 2.1273265713061515e+05, + "cpu_time": 3.1180410790273472e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.1613483217822766e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3290, + "real_time": 2.1284626575953281e+05, + "cpu_time": 3.1232124376899708e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.1580596461147761e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3290, + "real_time": 2.1260572561734472e+05, + "cpu_time": 3.1191605623100186e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.1650268175706613e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3290, + "real_time": 2.1265611036766978e+05, + "cpu_time": 3.1228659057750757e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.1635661337633002e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3290, + "real_time": 2.1271344853616756e+05, + "cpu_time": 3.1148233039513673e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.1619047080473554e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1271084148226600e+05, + "cpu_time": 3.1196206577507558e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.1619811254556739e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1271344853616756e+05, + "cpu_time": 3.1191605623100186e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 6.1619047080473554e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.0646696106365638e+01, + "cpu_time": 3.5054913387063277e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.6254891033862624e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.2614986370557409e-04, + "cpu_time": 1.1236915392250878e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2607873181242656e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 3045, + "real_time": 2.2850628429087112e+05, + "cpu_time": 3.4436548341543478e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1472069611280825e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 3045, + "real_time": 2.2887321296476794e+05, + "cpu_time": 3.4475671855500765e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1453677632443323e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 3045, + "real_time": 2.2864862753186157e+05, + "cpu_time": 3.4431183152709465e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1464927772788444e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 3045, + "real_time": 2.2899080783760324e+05, + "cpu_time": 3.4484806798029633e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1447795764182310e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 3045, + "real_time": 2.2846450622014594e+05, + "cpu_time": 3.4434171494252770e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1474167446710558e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2869668776904995e+05, + "cpu_time": 3.4452476328407222e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1462527645481093e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2864862753186157e+05, + "cpu_time": 3.4436548341543472e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 1.1464927772788444e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2920476761098848e+02, + "cpu_time": 2.5619615495054666e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1485270484929818e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0022216318342644e-03, + "cpu_time": 7.4362188804206314e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0019841033454512e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2722, + "real_time": 2.5676273750921359e+05, + "cpu_time": 4.1710060764144111e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0419162261860006e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2722, + "real_time": 2.5679789486156078e+05, + "cpu_time": 4.1639804775900196e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0416366741738386e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2722, + "real_time": 2.5709032340329018e+05, + "cpu_time": 4.1730592285084550e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0393144053794842e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2722, + "real_time": 2.5615454157013760e+05, + "cpu_time": 4.1573234790595254e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0467644133353961e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2722, + "real_time": 2.5665411548109070e+05, + "cpu_time": 4.1633322116091283e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0427804129196889e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5669192256505851e+05, + "cpu_time": 4.1657402946363075e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0424824263988817e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5676273750921356e+05, + "cpu_time": 4.1639804775900196e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 2.0419162261860006e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.4101899548973569e+02, + "cpu_time": 6.3449557742753836e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.7155314755826346e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.3285147116512965e-03, + "cpu_time": 1.5231280217936228e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.3295250135250424e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2250, + "real_time": 3.1092971431401867e+05, + "cpu_time": 5.6126341733333457e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.3723891661927390e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2250, + "real_time": 3.1175311640577595e+05, + "cpu_time": 5.6287484133333189e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.3634820145155497e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2250, + "real_time": 3.1159958172227361e+05, + "cpu_time": 5.6297028622222459e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.3651393054005704e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2250, + "real_time": 3.1133284238684503e+05, + "cpu_time": 5.6196984533333231e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.3680224417091761e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2250, + "real_time": 3.1084580041675101e+05, + "cpu_time": 5.6092989955555741e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.3732995542940388e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1129221104913286e+05, + "cpu_time": 5.6200165795555606e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.3684664964224148e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1133284238684509e+05, + "cpu_time": 5.6196984533333243e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 3.3680224417091761e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.9976484329034918e+02, + "cpu_time": 9.2132611728771121e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3260325669416776e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2842108767933554e-03, + "cpu_time": 1.6393654791683390e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2842735920147272e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1650, + "real_time": 4.2414169023025100e+05, + "cpu_time": 8.5175387030302954e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9444608919758224e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1650, + "real_time": 4.2369989006229764e+05, + "cpu_time": 8.5106394606060756e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9496165781200714e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1650, + "real_time": 4.2401505226735026e+05, + "cpu_time": 8.5199593696969619e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9459376236429043e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1650, + "real_time": 4.2441951696563401e+05, + "cpu_time": 8.5205768242423458e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9412242278429670e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1650, + "real_time": 4.2418930086899886e+05, + "cpu_time": 8.5193133151514735e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9439059299792604e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2409309007890633e+05, + "cpu_time": 8.5176055345454311e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9450290503122053e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2414169023025100e+05, + "cpu_time": 8.5193133151514735e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9444608919758224e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6404031714647749e+02, + "cpu_time": 4.0563613461978565e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0794396439871979e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.2259990394408543e-04, + "cpu_time": 4.7623259022106580e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.2273438895020787e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1049, + "real_time": 6.6710209472413862e+05, + "cpu_time": 1.4735922354623394e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.2873494674521093e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1049, + "real_time": 6.6753771658578410e+05, + "cpu_time": 1.4738162688274633e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.2832464680083694e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1049, + "real_time": 6.6657740809422894e+05, + "cpu_time": 1.4734970419447185e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.2922984623671541e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1049, + "real_time": 6.6760055772077700e+05, + "cpu_time": 1.4744449399428072e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.2826550270113201e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1049, + "real_time": 6.6714023677827534e+05, + "cpu_time": 1.4735581677788403e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.2869900041630688e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6719160278064082e+05, + "cpu_time": 1.4737817307912339e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.2865078858004045e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6714023677827534e+05, + "cpu_time": 1.4735922354623396e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 6.2869900041630688e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.1076884609702046e+02, + "cpu_time": 3.8986404355208731e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8714610163709512e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.1566848920919799e-04, + "cpu_time": 2.6453309564556738e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1583650043859489e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 605, + "real_time": 1.1556451197344163e+06, + "cpu_time": 2.7323427752066222e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2588096957721949e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 605, + "real_time": 1.1584476298413987e+06, + "cpu_time": 2.7274491669421587e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2412492234530039e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 605, + "real_time": 1.1547078601708098e+06, + "cpu_time": 2.7304117471074471e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2647015659520311e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 605, + "real_time": 1.1552336177411529e+06, + "cpu_time": 2.7323823669421296e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2613953326621351e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 605, + "real_time": 1.1533635553003342e+06, + "cpu_time": 2.7272626363636269e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2731689513248215e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1554795565576223e+06, + "cpu_time": 2.7299697385123973e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2598649538328371e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1552336177411529e+06, + "cpu_time": 2.7304117471074471e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 7.2613953326621351e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8688454135343327e+03, + "cpu_time": 2.5164028770687128e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1730198342398137e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6173764416065942e-03, + "cpu_time": 9.2176951325472932e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6157598546244572e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 321, + "real_time": 2.1728813079565857e+06, + "cpu_time": 5.3471768535825927e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.7211838210240660e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 321, + "real_time": 2.1747350634722696e+06, + "cpu_time": 5.3502539003115157e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.7146022436465530e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 321, + "real_time": 2.1830864890400218e+06, + "cpu_time": 5.3565408068535402e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.6850899330962925e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 321, + "real_time": 2.1746134685752825e+06, + "cpu_time": 5.3473497352025239e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.7150336105440121e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 321, + "real_time": 2.1730492978046336e+06, + "cpu_time": 5.3491981557632443e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.7205869268357210e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1756731253697588e+06, + "cpu_time": 5.3501038903426835e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.7112993070293312e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.1746134685752830e+06, + "cpu_time": 5.3491981557632443e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 7.7150336105440121e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2320062946624075e+03, + "cpu_time": 3.8219650378254469e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4964534685726784e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.9451480304253754e-03, + "cpu_time": 7.1437211616102721e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.9405983466477142e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 163, + "real_time": 4.2048094342776965e+06, + "cpu_time": 1.0809165226993905e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.9800125367070265e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 163, + "real_time": 4.1902699890586496e+06, + "cpu_time": 1.0771673128834350e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 8.0077016725927134e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 163, + "real_time": 4.2086896983666653e+06, + "cpu_time": 1.0800486987730080e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.9726552454133205e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 163, + "real_time": 4.2332571822624262e+06, + "cpu_time": 1.0855465625766844e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.9263863628684931e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 163, + "real_time": 4.1801162875411701e+06, + "cpu_time": 1.0756251815950919e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 8.0271527612781801e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2034285183013221e+06, + "cpu_time": 1.0798608557055222e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.9827817157719469e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2048094342776965e+06, + "cpu_time": 1.0800486987730080e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 7.9800125367070265e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.0223396927562440e+04, + "cpu_time": 3.8299180326467547e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.8342024956136055e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.8111670840866507e-03, + "cpu_time": 3.5466773449663530e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.8030907422135774e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 77, + "real_time": 8.9882637749051116e+06, + "cpu_time": 2.3019242727272812e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.4662766559394236e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 77, + "real_time": 8.9990441481788438e+06, + "cpu_time": 2.3022153376623258e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.4573324560899019e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 77, + "real_time": 8.9861670170318000e+06, + "cpu_time": 2.2993756610389754e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.4680187751692352e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 77, + "real_time": 8.9883780372994281e+06, + "cpu_time": 2.3032328454545509e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.4661817428590212e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 77, + "real_time": 8.9825431199429873e+06, + "cpu_time": 2.2970112779220782e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.4710316559466677e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9888792194716334e+06, + "cpu_time": 2.3007518789610423e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.4657682572008505e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.9882637749051116e+06, + "cpu_time": 2.3019242727272809e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 7.4662766559394236e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.1534122749181206e+03, + "cpu_time": 2.5277205717316076e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1078463550893934e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.8455834422478956e-04, + "cpu_time": 1.0986497913338911e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.8416888645891135e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 41, + "real_time": 1.6993939649404548e+07, + "cpu_time": 4.3404790073170871e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.8979760296314135e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 41, + "real_time": 1.7013927921652794e+07, + "cpu_time": 4.3434386048780546e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.8886973436150312e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 41, + "real_time": 1.7009923688885644e+07, + "cpu_time": 4.3432133097560838e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.8905543878305845e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 41, + "real_time": 1.7003844015118554e+07, + "cpu_time": 4.3413431414633788e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.8933756320431776e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 41, + "real_time": 1.7002989060995057e+07, + "cpu_time": 4.3393068829268150e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.8937725313189878e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7004924867211320e+07, + "cpu_time": 4.3415561892682835e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.8928751848878393e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7003844015118551e+07, + "cpu_time": 4.3413431414633796e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 7.8933756320431776e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6097602531783141e+03, + "cpu_time": 1.7716253968153160e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.5325186395205334e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.4750331522201296e-04, + "cpu_time": 4.0806229830550730e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.4755789959584059e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 21, + "real_time": 3.2920341406549726e+07, + "cpu_time": 8.3824041047618955e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.1540908912503843e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 21, + "real_time": 3.2898314829383578e+07, + "cpu_time": 8.3822344571428537e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.1595503414735155e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 21, + "real_time": 3.2950390839860551e+07, + "cpu_time": 8.3804190000000477e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.1466546877880983e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 21, + "real_time": 3.2877285920438312e+07, + "cpu_time": 8.3823741428570241e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.1647693380044460e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 21, + "real_time": 3.3011015859388169e+07, + "cpu_time": 8.3963997333332449e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.1316932851570606e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2931469771124065e+07, + "cpu_time": 8.3847662876190141e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.1513517087347002e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.2920341406549722e+07, + "cpu_time": 8.3823741428570241e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 8.1540908912503843e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.2068626193492179e+04, + "cpu_time": 6.5564548124357636e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2877005652035724e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5811206288505381e-03, + "cpu_time": 7.8194842736607425e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5797386877856321e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10, + "real_time": 6.5180983394384384e+07, + "cpu_time": 1.6464316130000043e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.2366187811498051e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10, + "real_time": 6.5075767785310745e+07, + "cpu_time": 1.6456440960000122e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.2499358865987196e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10, + "real_time": 6.4986293762922287e+07, + "cpu_time": 1.6435073590000114e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.2612945117099428e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10, + "real_time": 6.4945159107446671e+07, + "cpu_time": 1.6435911149999923e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.2665270110708208e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10, + "real_time": 6.5094535797834396e+07, + "cpu_time": 1.6451428360000193e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.2475572706651201e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.5056547969579712e+07, + "cpu_time": 1.6448634038000080e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.2523866922388811e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.5075767785310745e+07, + "cpu_time": 1.6451428360000193e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 8.2499358865987196e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.3058830205128732e+04, + "cpu_time": 1.2849532140731267e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1803120832593387e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4304298815338747e-03, + "cpu_time": 7.8119144185747767e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4302675423212856e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 5, + "real_time": 1.4154558479785919e+08, + "cpu_time": 3.1850741679999715e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.5858376333914433e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 5, + "real_time": 1.4167163670063019e+08, + "cpu_time": 3.1851921479999983e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.5790881577019558e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 5, + "real_time": 1.4156373143196106e+08, + "cpu_time": 3.1830709879999971e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.5848652274051294e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 5, + "real_time": 1.4166257381439209e+08, + "cpu_time": 3.1886318599999869e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.5795730311015577e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 5, + "real_time": 1.4162274301052094e+08, + "cpu_time": 3.1861687819999814e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.5817047543008919e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4161325395107269e+08, + "cpu_time": 3.1856275891999876e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.5822137607801962e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4162274301052094e+08, + "cpu_time": 3.1851921479999983e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 7.5817047543008919e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.6926774924425314e+04, + "cpu_time": 2.0217953935040539e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.0481413745690994e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_GPUToHost/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.0198762005774888e-04, + "cpu_time": 6.3466156570165546e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0201206016321164e-04, + "hip_id": NaN, + "numa_id": NaN + } + ] +} diff --git a/results/prefetch_managed_HostToGPU.json b/results/prefetch_managed_HostToGPU.json new file mode 100644 index 0000000..656fe6c --- /dev/null +++ b/results/prefetch_managed_HostToGPU.json @@ -0,0 +1,3526 @@ +{ + "context": { + "date": "2023-10-09T17:47:41-04:00", + "host_name": "frontier08051", + "executable": "/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth/build/comm_scope", + "num_cpus": 128, + "mhz_per_cpu": 1745, + "cpu_scaling_enabled": false, + "caches": [ + { + "type": "Data", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Instruction", + "level": 1, + "size": 32768, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 2, + "size": 524288, + "num_sharing": 2 + }, + { + "type": "Unified", + "level": 3, + "size": 33554432, + "num_sharing": 16 + } + ], + "load_avg": [1.9,3,15.5], + "library_build_type": "release" + }, + "benchmarks": [ + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9239, + "real_time": 7.6772328023020062e+04, + "cpu_time": 2.7883222285961692e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.3338139227625810e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9239, + "real_time": 7.6584782229995471e+04, + "cpu_time": 2.7845749312696181e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.3370802529995792e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9239, + "real_time": 7.6530334371943129e+04, + "cpu_time": 2.7825278947938071e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.3380315248895735e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9239, + "real_time": 7.6379357563150625e+04, + "cpu_time": 2.7852738748782332e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.3406763720856836e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9239, + "real_time": 7.6531916843503801e+04, + "cpu_time": 2.7851779229353816e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.3380038580425540e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6559743806322615e+04, + "cpu_time": 2.7851753704946412e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.3375211861559942e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time_median", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6531916843503801e+04, + "cpu_time": 2.7851779229353816e+05, + "time_unit": "ns", + "bytes": 1.0240000000000000e+03, + "bytes_per_second": 1.3380038580425540e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4141907867909799e+02, + "cpu_time": 2.0785860816542399e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.4691516731777028e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 0, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):10/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.8471728306308545e-03, + "cpu_time": 7.4630348367797292e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.8460654670256020e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9042, + "real_time": 7.6799245522813508e+04, + "cpu_time": 2.7859990864852915e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.6666928640485588e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9042, + "real_time": 7.6705982475950179e+04, + "cpu_time": 2.7865494359654927e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.6699351652814232e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9042, + "real_time": 7.6635171049932149e+04, + "cpu_time": 2.7820108250387060e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.6724022037683088e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9042, + "real_time": 7.6804080461827572e+04, + "cpu_time": 2.7858089593010413e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.6665249914916661e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9042, + "real_time": 7.6510161350394847e+04, + "cpu_time": 2.7851856359212578e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.6767686328888271e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6690928172183660e+04, + "cpu_time": 2.7851107885423588e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.6704647714957569e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time_median", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6705982475950179e+04, + "cpu_time": 2.7858089593010419e+05, + "time_unit": "ns", + "bytes": 2.0480000000000000e+03, + "bytes_per_second": 2.6699351652814232e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2303274126955944e+02, + "cpu_time": 1.8001109619643603e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.2872469980017479e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 1, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):11/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6042672086759833e-03, + "cpu_time": 6.4633370039347089e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6054310260009209e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9261, + "real_time": 7.6404890036560886e+04, + "cpu_time": 2.7828459313249099e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.3609134154109806e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9261, + "real_time": 7.6599688664451896e+04, + "cpu_time": 2.7846538375985355e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.3472802193004943e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9261, + "real_time": 7.6695179677223045e+04, + "cpu_time": 2.7836910322859342e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.3406224709796615e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9261, + "real_time": 7.6731064343488499e+04, + "cpu_time": 2.7846796317892213e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.3381248325504199e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9261, + "real_time": 7.6550964385814572e+04, + "cpu_time": 2.7827732998596231e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.3506837345069654e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6596357421507797e+04, + "cpu_time": 2.7837287465716444e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.3475249345497042e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time_median", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.6599688664451911e+04, + "cpu_time": 2.7836910322859342e+05, + "time_unit": "ns", + "bytes": 4.0960000000000000e+03, + "bytes_per_second": 5.3472802193004943e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2907669950553773e+02, + "cpu_time": 9.2919853194684293e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.0179715488711765e+04, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 2, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):12/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.6851545406426050e-03, + "cpu_time": 3.3379636327397750e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.6863823281322479e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 9178, + "real_time": 7.7364956940136195e+04, + "cpu_time": 2.8018909424711327e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 1.0588773424043709e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 9178, + "real_time": 7.7558495322163537e+04, + "cpu_time": 2.8049972837219416e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 1.0562350347272673e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 9178, + "real_time": 7.7301074610206211e+04, + "cpu_time": 2.8002235334495484e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 1.0597524085284054e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 9178, + "real_time": 7.7400826057138867e+04, + "cpu_time": 2.8031483155371575e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 1.0583866371080458e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 9178, + "real_time": 7.7276126120525048e+04, + "cpu_time": 2.8010548115057760e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 1.0600945481173845e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7380295810033975e+04, + "cpu_time": 2.8022629773371120e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 1.0586691941770948e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time_median", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.7364956940136180e+04, + "cpu_time": 2.8018909424711327e+05, + "time_unit": "ns", + "bytes": 8.1920000000000000e+03, + "bytes_per_second": 1.0588773424043709e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1128697229564682e+02, + "cpu_time": 1.8720083545871222e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5209605184717977e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 3, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):13/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4381823063697326e-03, + "cpu_time": 6.6803450273108329e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4366721227342810e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8980, + "real_time": 7.8648558167440206e+04, + "cpu_time": 2.8280538151447632e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 2.0831914000405452e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8980, + "real_time": 7.8579033917962894e+04, + "cpu_time": 2.8262714298440929e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 2.0850345420516393e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8980, + "real_time": 7.8843107970864701e+04, + "cpu_time": 2.8296967906458810e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 2.0780510080924848e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8980, + "real_time": 7.8925282641020734e+04, + "cpu_time": 2.8306425957683724e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 2.0758874028389680e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8980, + "real_time": 7.8944591091193084e+04, + "cpu_time": 2.8285137349665922e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 2.0753796775100112e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8788114757696327e+04, + "cpu_time": 2.8286356732739409e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 2.0795088061067298e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time_median", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.8843107970864701e+04, + "cpu_time": 2.8285137349665922e+05, + "time_unit": "ns", + "bytes": 1.6384000000000000e+04, + "bytes_per_second": 2.0780510080924848e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.6546522601169551e+02, + "cpu_time": 1.6655916967397326e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.3699169174024352e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 4, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):14/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 2.1001292710272935e-03, + "cpu_time": 5.8883217534053470e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.1014178466422667e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8510, + "real_time": 8.1663406601768816e+04, + "cpu_time": 2.8788041280846030e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 4.0125683416310298e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8510, + "real_time": 8.1433502763792101e+04, + "cpu_time": 2.8766524935370084e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 4.0238966626607746e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8510, + "real_time": 8.1429984638218186e+04, + "cpu_time": 2.8768611903642758e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 4.0240705123037362e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8510, + "real_time": 8.1539279531796114e+04, + "cpu_time": 2.8753462996474718e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 4.0186766657929778e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8510, + "real_time": 8.1334735485775716e+04, + "cpu_time": 2.8757357732079935e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 4.0287830044926691e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.1480181804270193e+04, + "cpu_time": 2.8766799769682711e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 4.0215990373762375e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time_median", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.1433502763792101e+04, + "cpu_time": 2.8766524935370084e+05, + "time_unit": "ns", + "bytes": 3.2768000000000000e+04, + "bytes_per_second": 4.0238966626607746e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2541649380740401e+02, + "cpu_time": 1.3431146189960506e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.1864847672971769e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 5, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):15/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5392269755690603e-03, + "cpu_time": 4.6689747547502917e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5383146628494692e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8191, + "real_time": 8.6326750542876645e+04, + "cpu_time": 2.9606204541570146e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.5916213210700750e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8191, + "real_time": 8.6293681920443516e+04, + "cpu_time": 2.9603306580393057e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.5945305080874181e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8191, + "real_time": 8.6334783706406204e+04, + "cpu_time": 2.9594632755463279e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.5909149460389626e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8191, + "real_time": 8.6250047906696273e+04, + "cpu_time": 2.9575643767549715e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.5983725911544585e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8191, + "real_time": 8.6275951978675774e+04, + "cpu_time": 2.9604854059333442e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.5960912046728945e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6296243211019682e+04, + "cpu_time": 2.9596928340861935e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.5943061142047632e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time_median", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.6293681920443516e+04, + "cpu_time": 2.9603306580393051e+05, + "time_unit": "ns", + "bytes": 6.5536000000000000e+04, + "bytes_per_second": 7.5945305080874181e+08, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.5243323972834908e+01, + "cpu_time": 1.2726445160999842e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.1016637058198301e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 6, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):16/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 4.0839928438894630e-04, + "cpu_time": 4.2999209291018001e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.0841963165249895e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 7191, + "real_time": 9.4308981830678997e+04, + "cpu_time": 3.1162975928243803e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.3898146014907126e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 7191, + "real_time": 9.4474132929098691e+04, + "cpu_time": 3.1175028897232545e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.3873850538365610e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 7191, + "real_time": 9.7115343972771399e+04, + "cpu_time": 3.1158989751077612e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.3496528420550017e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 7191, + "real_time": 9.4510706860076447e+04, + "cpu_time": 3.1175154456960142e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.3868481609607756e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 7191, + "real_time": 9.4445115155483421e+04, + "cpu_time": 3.1173838589904120e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.3878113207254641e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.4970856149621788e+04, + "cpu_time": 3.1169197524683643e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.3803023958137031e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time_median", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.4474132929098691e+04, + "cpu_time": 3.1173838589904114e+05, + "time_unit": "ns", + "bytes": 1.3107200000000000e+05, + "bytes_per_second": 1.3873850538365610e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.2012265145582737e+03, + "cpu_time": 7.6474887795034206e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7170266755416468e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 7, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):17/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2648369860600204e-02, + "cpu_time": 2.4535404780463756e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2439496452003485e-02, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 6088, + "real_time": 1.1505584992698510e+05, + "cpu_time": 3.4595669645203697e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2784065318396034e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 6088, + "real_time": 1.1509140589451614e+05, + "cpu_time": 3.4579126511169429e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2777026482781944e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 6088, + "real_time": 1.1493658654699975e+05, + "cpu_time": 3.4555477398160403e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2807707090970926e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 6088, + "real_time": 1.1504483297983147e+05, + "cpu_time": 3.4550262828514935e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2786247170783978e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 6088, + "real_time": 1.1513671190339846e+05, + "cpu_time": 3.4597141409329831e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2768063779686794e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1505307745034617e+05, + "cpu_time": 3.4575535558475653e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2784621968523936e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time_median", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.1505584992698510e+05, + "cpu_time": 3.4579126511169423e+05, + "time_unit": "ns", + "bytes": 2.6214400000000000e+05, + "bytes_per_second": 2.2784065318396034e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.4346452220684156e+01, + "cpu_time": 2.1943847982246967e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4728705317168920e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 8, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):18/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 6.4619264315437445e-04, + "cpu_time": 6.3466400817232670e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.4643184940772996e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4419, + "real_time": 1.5943640453586471e+05, + "cpu_time": 4.1796502489250881e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.2883832367284918e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4419, + "real_time": 1.5928610465627370e+05, + "cpu_time": 4.1784324869880168e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.2914861037713885e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4419, + "real_time": 1.5936803732970532e+05, + "cpu_time": 4.1757537044580100e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.2897939184338288e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4419, + "real_time": 1.5932849623601881e+05, + "cpu_time": 4.1751149490834924e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.2906103577564311e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4419, + "real_time": 1.5984062332649736e+05, + "cpu_time": 4.1854947974654782e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.2800672888334942e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5945193321687198e+05, + "cpu_time": 4.1788892373840173e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.2880681811047268e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time_median", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5936803732970537e+05, + "cpu_time": 4.1784324869880168e+05, + "time_unit": "ns", + "bytes": 5.2428800000000000e+05, + "bytes_per_second": 3.2897939184338288e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.2421834273353548e+02, + "cpu_time": 4.1381759985448014e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 4.6161825136881229e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 9, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):19/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.4061814003131221e-03, + "cpu_time": 9.9025740178155510e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.4039193409113480e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 2822, + "real_time": 2.5061075746848743e+05, + "cpu_time": 5.6412251381998602e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.1840821622824841e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 2822, + "real_time": 2.4786720431350704e+05, + "cpu_time": 5.6416356378454622e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.2303942665756688e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 2822, + "real_time": 2.5053216500825252e+05, + "cpu_time": 5.6420785683911503e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.1853947175423956e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 2822, + "real_time": 2.4784452208544873e+05, + "cpu_time": 5.6403418284904421e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.2307814236801457e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 2822, + "real_time": 2.5055455794093589e+05, + "cpu_time": 5.6402367044649681e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.1850206542528138e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.4948184136332633e+05, + "cpu_time": 5.6411035754783778e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.2031346448667016e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time_median", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5053216500825254e+05, + "cpu_time": 5.6412251381998602e+05, + "time_unit": "ns", + "bytes": 1.0485760000000000e+06, + "bytes_per_second": 4.1853947175423956e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4846059584770105e+03, + "cpu_time": 8.0315049378503630e+01, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.5066163071627855e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 10, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):20/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.9507575796466233e-03, + "cpu_time": 1.4237471144410384e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.9636831054749153e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 1637, + "real_time": 4.2786900269546901e+05, + "cpu_time": 8.5447645510080724e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9013880107894249e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 1637, + "real_time": 4.2812537673787394e+05, + "cpu_time": 8.5469839890043973e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.8984529157775478e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 1637, + "real_time": 4.2848848688280210e+05, + "cpu_time": 8.5519725229078264e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.8943018638762207e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 1637, + "real_time": 4.2788699776985019e+05, + "cpu_time": 8.5412804459376691e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.9011818796326370e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 1637, + "real_time": 4.2814081591990916e+05, + "cpu_time": 8.5431845021380601e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.8982762727118902e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2810213600118097e+05, + "cpu_time": 8.5456372021992062e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.8987201885575438e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time_median", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.2812537673787400e+05, + "cpu_time": 8.5447645510080713e+05, + "time_unit": "ns", + "bytes": 2.0971520000000000e+06, + "bytes_per_second": 4.8984529157775478e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.5096756174232638e+02, + "cpu_time": 4.1142650015963500e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8708573388170996e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 11, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):21/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.8623291181531055e-04, + "cpu_time": 4.8144625195854920e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8604231887399150e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 867, + "real_time": 8.0706062431274261e+05, + "cpu_time": 1.4754427289503908e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 5.1970123106571884e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 867, + "real_time": 8.0719294502897223e+05, + "cpu_time": 1.4760323817762402e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 5.1961603800308933e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 867, + "real_time": 8.0695672007975832e+05, + "cpu_time": 1.4752122479815455e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 5.1976814810903883e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 867, + "real_time": 8.0720587646171905e+05, + "cpu_time": 1.4761695109573239e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 5.1960771375763283e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 867, + "real_time": 8.0733393767962477e+05, + "cpu_time": 1.4756545570934466e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 5.1952529235361223e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0715002071256354e+05, + "cpu_time": 1.4757022853517896e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 5.1964368465781841e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time_median", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0719294502897211e+05, + "cpu_time": 1.4756545570934464e+06, + "time_unit": "ns", + "bytes": 4.1943040000000000e+06, + "bytes_per_second": 5.1961603800308933e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4503849803819747e+02, + "cpu_time": 3.9907629286790586e+02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 9.3377889220093214e+05, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 12, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):22/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.7969211957667477e-04, + "cpu_time": 2.7043143920643241e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.7969599550042039e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 444, + "real_time": 1.5759804896456567e+06, + "cpu_time": 2.7297354481981909e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.3227867065068130e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 444, + "real_time": 1.5771577747219855e+06, + "cpu_time": 2.7351972500000373e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.3188134595340071e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 444, + "real_time": 1.5800345774534419e+06, + "cpu_time": 2.7368444864864773e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.3091293821683359e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 444, + "real_time": 1.5777322067249869e+06, + "cpu_time": 2.7371065292793070e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.3168769479662466e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 444, + "real_time": 1.5803603292862372e+06, + "cpu_time": 2.7369406328828675e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.3080350376731348e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5782530755664618e+06, + "cpu_time": 2.7351648693693755e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.3151283067697077e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time_median", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.5777322067249869e+06, + "cpu_time": 2.7368444864864773e+06, + "time_unit": "ns", + "bytes": 8.3886080000000000e+06, + "bytes_per_second": 5.3168769479662466e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.8874521365197229e+03, + "cpu_time": 3.1314595511800808e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.3559174459081832e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 13, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):23/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.1959122182241177e-03, + "cpu_time": 1.1448887729762613e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.1958163715093870e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 220, + "real_time": 3.1731148165735332e+06, + "cpu_time": 5.3441974954545591e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.2873019004451799e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 220, + "real_time": 3.1822283053770661e+06, + "cpu_time": 5.3641568818182927e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.2721597541104288e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 220, + "real_time": 3.1726050000129775e+06, + "cpu_time": 5.3451815272726901e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.2881515347581472e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 220, + "real_time": 3.1730603024532851e+06, + "cpu_time": 5.3467545181817403e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.2873927378652458e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 220, + "real_time": 3.1737635875205426e+06, + "cpu_time": 5.3460940090909572e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.2862210865261583e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1749544023874807e+06, + "cpu_time": 5.3492768863636479e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.2842454027410316e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time_median", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 3.1731148165735332e+06, + "cpu_time": 5.3460940090909563e+06, + "time_unit": "ns", + "bytes": 1.6777216000000000e+07, + "bytes_per_second": 5.2873019004451799e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 4.0871471342478717e+03, + "cpu_time": 8.3737609798413141e+03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 6.7910233617739826e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 14, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):24/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2873089236098781e-03, + "cpu_time": 1.5654005499673549e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2851453413294089e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 102, + "real_time": 6.7100741891373023e+06, + "cpu_time": 1.0978727078431368e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0006052174982004e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 102, + "real_time": 6.5903511884457925e+06, + "cpu_time": 1.0749629372549064e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0914482461613960e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 102, + "real_time": 6.7902221527023641e+06, + "cpu_time": 1.1144645509803826e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 4.9415808857808943e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 102, + "real_time": 6.6938718242169013e+06, + "cpu_time": 1.0958121166666780e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0127090690036402e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 102, + "real_time": 6.6548816667979257e+06, + "cpu_time": 1.0872203372548928e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0420779331670837e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6878802042600578e+06, + "cpu_time": 1.0940665299999993e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0176842703222427e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time_median", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 6.6938718242169013e+06, + "cpu_time": 1.0958121166666780e+07, + "time_unit": "ns", + "bytes": 3.3554432000000000e+07, + "bytes_per_second": 5.0127090690036402e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 7.3490066141107600e+04, + "cpu_time": 1.4538147521047213e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.5100695845792726e+07, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 15, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):25/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.0988544037361162e-02, + "cpu_time": 1.3288175008010912e-02, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.0981299913925050e-02, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 51, + "real_time": 1.4027418185244588e+07, + "cpu_time": 2.2831177725490246e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7841208634238672e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 51, + "real_time": 1.4021498205907205e+07, + "cpu_time": 2.2821028901961092e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7861407543259029e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 51, + "real_time": 1.4020208780672036e+07, + "cpu_time": 2.2810221372549001e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7865809311281347e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 51, + "real_time": 1.3984701157931019e+07, + "cpu_time": 2.2771878058823545e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7987342197828197e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 51, + "real_time": 1.4006903687236356e+07, + "cpu_time": 2.2796106058823332e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7911276823551121e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4012146003398240e+07, + "cpu_time": 2.2806082423529442e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7893408902031679e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time_median", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4020208780672034e+07, + "cpu_time": 2.2810221372549001e+07, + "time_unit": "ns", + "bytes": 6.7108864000000000e+07, + "bytes_per_second": 4.7865809311281347e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7076394738077204e+04, + "cpu_time": 2.3128817063733295e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.8423922771412740e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 16, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):26/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.2186851845488776e-03, + "cpu_time": 1.0141512529074647e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.2198739682723722e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 27, + "real_time": 2.6351949269021001e+07, + "cpu_time": 4.2719197999999821e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0932751361124001e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 27, + "real_time": 2.6401039764836982e+07, + "cpu_time": 4.2821852037037529e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0838046226786079e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 27, + "real_time": 2.6381697643686224e+07, + "cpu_time": 4.2738681222221814e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0875318871726036e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 27, + "real_time": 2.6398586040293727e+07, + "cpu_time": 4.2747790148147851e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0842771576907759e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 27, + "real_time": 2.6385703641507361e+07, + "cpu_time": 4.2732625814814441e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0867594748870764e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6383795271869060e+07, + "cpu_time": 4.2752029444444291e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0871296557082930e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time_median", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 2.6385703641507357e+07, + "cpu_time": 4.2738681222221814e+07, + "time_unit": "ns", + "bytes": 1.3421772800000000e+08, + "bytes_per_second": 5.0867594748870764e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.9610897376185549e+04, + "cpu_time": 4.0391043849298396e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 3.7834621378414771e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 17, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):27/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 7.4329326672327109e-04, + "cpu_time": 9.4477488844795150e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 7.4373220143819922e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 10, + "real_time": 5.0603139400482178e+07, + "cpu_time": 8.0306016000000119e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3047194142551994e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 10, + "real_time": 5.0435474887490273e+07, + "cpu_time": 8.0080426899999678e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3223540890378571e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 10, + "real_time": 5.0612307712435722e+07, + "cpu_time": 8.0219879400002018e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3037584756097565e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 10, + "real_time": 5.0526370480656624e+07, + "cpu_time": 8.0143898000000030e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3127793159567451e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 10, + "real_time": 5.0460274145007133e+07, + "cpu_time": 8.0097571999999672e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3197383594984045e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0527513325214393e+07, + "cpu_time": 8.0169558460000306e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3126699308715935e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time_median", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.0526370480656624e+07, + "cpu_time": 8.0143898000000030e+07, + "time_unit": "ns", + "bytes": 2.6843545600000000e+08, + "bytes_per_second": 5.3127793159567451e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 8.0471856618870923e+04, + "cpu_time": 9.3458281916243839e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4614170951962881e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 18, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):28/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 1.5926344148567790e-03, + "cpu_time": 1.1657577229001927e-03, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 1.5926863903265515e-03, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 8, + "real_time": 9.9050653167068958e+07, + "cpu_time": 1.5603962112500015e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4201652874964752e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 8, + "real_time": 9.9029911682009697e+07, + "cpu_time": 1.5594765362499884e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4213005230573263e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 8, + "real_time": 9.8958233371376991e+07, + "cpu_time": 1.5604148087500036e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4252273278282509e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 8, + "real_time": 9.9049472250044346e+07, + "cpu_time": 1.5614531062499991e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4202299094002457e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 8, + "real_time": 9.8944511264562607e+07, + "cpu_time": 1.5601335075000122e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4259797247821932e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.9006556347012520e+07, + "cpu_time": 1.5603748340000013e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4225805545128984e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time_median", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 9.9029911682009697e+07, + "cpu_time": 1.5603962112500015e+08, + "time_unit": "ns", + "bytes": 5.3687091200000000e+08, + "bytes_per_second": 5.4213005230573263e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 5.1274940677028855e+04, + "cpu_time": 7.1237338128141186e+04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 2.8087940128674442e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 19, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):29/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 5.1789439577428604e-04, + "cpu_time": 4.5653990679614571e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1798105802777761e-04, + "hip_id": NaN, + "numa_id": NaN + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 0, + "threads": 1, + "iterations": 4, + "real_time": 1.7600608617067337e+08, + "cpu_time": 2.8251493700000197e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.1005948564687185e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 1, + "threads": 1, + "iterations": 4, + "real_time": 1.7639649286866188e+08, + "cpu_time": 2.8285736624999911e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.0870928131177034e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 2, + "threads": 1, + "iterations": 4, + "real_time": 1.7614972218871117e+08, + "cpu_time": 2.8245364450000209e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.0956203090101290e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 3, + "threads": 1, + "iterations": 4, + "real_time": 1.7630544677376747e+08, + "cpu_time": 2.8272291950000292e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.0902362555923157e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "run_type": "iteration", + "repetitions": 5, + "repetition_index": 4, + "threads": 1, + "iterations": 4, + "real_time": 1.7619436606764793e+08, + "cpu_time": 2.8294378374999952e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.0940758093692303e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time_mean", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "mean", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7621042281389236e+08, + "cpu_time": 2.8269853020000112e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.0935240087116203e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time_median", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "median", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.7619436606764793e+08, + "cpu_time": 2.8272291950000298e+08, + "time_unit": "ns", + "bytes": 1.0737418240000000e+09, + "bytes_per_second": 6.0940758093692303e+09, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time_stddev", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "stddev", + "aggregate_unit": "time", + "iterations": 5, + "real_time": 1.4945523324482297e+05, + "cpu_time": 2.1192481082279095e+05, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 5.1687871935455035e+06, + "hip_id": 0.0000000000000000e+00, + "numa_id": 0.0000000000000000e+00 + }, + { + "name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time_cv", + "family_index": 0, + "per_family_instance_index": 20, + "run_name": "Comm_prefetch_managed_HostToGPU/0/0/log2(N):30/manual_time", + "run_type": "aggregate", + "repetitions": 5, + "threads": 1, + "aggregate_name": "cv", + "aggregate_unit": "percentage", + "iterations": 5, + "real_time": 8.4816341087083516e-04, + "cpu_time": 7.4964949648963588e-04, + "time_unit": "ns", + "bytes": 0.0000000000000000e+00, + "bytes_per_second": 8.4824268947753965e-04, + "hip_id": NaN, + "numa_id": NaN + } + ] +} diff --git a/results/rocm-smi.r1.1465528.txt b/results/rocm-smi.r1.1465528.txt new file mode 100644 index 0000000..5b70240 --- /dev/null +++ b/results/rocm-smi.r1.1465528.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 38.0c 94.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 41.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 36.0c 91.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 34.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 30.0c 88.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 39.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 30.0c 88.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 39.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r10.1465527.txt b/results/rocm-smi.r10.1465527.txt new file mode 100644 index 0000000..87db748 --- /dev/null +++ b/results/rocm-smi.r10.1465527.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 36.0c 97.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 39.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 31.0c 89.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 42.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 33.0c 87.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 38.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 25.0c 93.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 33.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r11.1465787.txt b/results/rocm-smi.r11.1465787.txt new file mode 100644 index 0000000..d8372b9 --- /dev/null +++ b/results/rocm-smi.r11.1465787.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 31.0c 94.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 39.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 36.0c 85.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 44.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 36.0c 85.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 37.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 32.0c 88.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 41.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r12.1467862.txt b/results/rocm-smi.r12.1467862.txt new file mode 100644 index 0000000..375646f --- /dev/null +++ b/results/rocm-smi.r12.1467862.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 38.0c 98.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 34.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 31.0c 91.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 34.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 30.0c 94.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 37.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 29.0c 92.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 31.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r2.1465529.txt b/results/rocm-smi.r2.1465529.txt new file mode 100644 index 0000000..352ca09 --- /dev/null +++ b/results/rocm-smi.r2.1465529.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 36.0c 91.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 33.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 31.0c 88.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 39.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 38.0c 87.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 36.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 26.0c 86.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 41.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r3.1465530.txt b/results/rocm-smi.r3.1465530.txt new file mode 100644 index 0000000..a7c0fdf --- /dev/null +++ b/results/rocm-smi.r3.1465530.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 36.0c 86.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 40.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 31.0c 89.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 36.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 45.0c 85.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 42.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 32.0c 90.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 38.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r4.1465531.txt b/results/rocm-smi.r4.1465531.txt new file mode 100644 index 0000000..28efb5a --- /dev/null +++ b/results/rocm-smi.r4.1465531.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 40.0c 91.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 36.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 29.0c 92.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 35.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 32.0c 86.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 37.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 30.0c 85.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 37.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r5.1465532.txt b/results/rocm-smi.r5.1465532.txt new file mode 100644 index 0000000..2ebb443 --- /dev/null +++ b/results/rocm-smi.r5.1465532.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 37.0c 94.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 38.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 33.0c 86.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 36.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 39.0c 83.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 39.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 34.0c 85.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 33.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r5.1465786.txt b/results/rocm-smi.r5.1465786.txt new file mode 100644 index 0000000..7f8a156 --- /dev/null +++ b/results/rocm-smi.r5.1465786.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 40.0c 97.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 37.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 31.0c 90.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 33.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 40.0c 98.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 38.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 33.0c 92.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 36.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r6.1465533.txt b/results/rocm-smi.r6.1465533.txt new file mode 100644 index 0000000..07961df --- /dev/null +++ b/results/rocm-smi.r6.1465533.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 37.0c 91.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 42.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 35.0c 88.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 41.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 34.0c 92.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 35.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 32.0c 88.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 34.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r7.1465534.txt b/results/rocm-smi.r7.1465534.txt new file mode 100644 index 0000000..fe40172 --- /dev/null +++ b/results/rocm-smi.r7.1465534.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 36.0c 92.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 43.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 30.0c 88.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 34.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 32.0c 93.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 38.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 28.0c 92.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 36.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r8.1465535.txt b/results/rocm-smi.r8.1465535.txt new file mode 100644 index 0000000..0b49b75 --- /dev/null +++ b/results/rocm-smi.r8.1465535.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 42.0c 94.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 34.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 32.0c 86.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 37.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 32.0c 93.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 37.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 31.0c 84.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 40.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/results/rocm-smi.r9.1465536.txt b/results/rocm-smi.r9.1465536.txt new file mode 100644 index 0000000..770a02c --- /dev/null +++ b/results/rocm-smi.r9.1465536.txt @@ -0,0 +1,15 @@ + + +======================= ROCm System Management Interface ======================= +================================= Concise Info ================================= +GPU Temp AvgPwr SCLK MCLK Fan Perf PwrCap VRAM% GPU% +0 34.0c 96.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +1 36.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +2 32.0c 88.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +3 35.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +4 29.0c 89.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +5 39.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +6 30.0c 93.0W 800Mhz 1600Mhz 0% manual 560.0W 0% 0% +7 37.0c N/A 800Mhz 1600Mhz 0% manual 0.0W 0% 0% +================================================================================ +============================= End of ROCm SMI Log ============================== diff --git a/scripts/run_1.sh b/scripts/run_1.sh index a3a5d85..0b99b9a 100755 --- a/scripts/run_1.sh +++ b/scripts/run_1.sh @@ -35,7 +35,7 @@ srun -c 56 -n 1 --gpus 8 $SCOPE_BUILD/comm_scope \ --benchmark_repetitions=5 \ --benchmark_filter='.*hipMemcpyAsync_PinnedToGPU/0/0/.*' \ --benchmark_out_format=json \ ---benchmark_out="$SCOPE_RESULTS/hipMemcpyAsync_PinnedToGPU" +--benchmark_out="$SCOPE_RESULTS/hipMemcpyAsync_PinnedToGPU.json" srun -c 56 -n 1 --gpus 8 $SCOPE_BUILD/comm_scope \ --benchmark_repetitions=5 \ diff --git a/scripts/run_11.sh b/scripts/run_11.sh index e994b6d..d495a3e 100755 --- a/scripts/run_11.sh +++ b/scripts/run_11.sh @@ -31,22 +31,4 @@ srun -c 56 -n 1 --gpus 8 $SCOPE_BUILD/comm_scope \ --benchmark_out_format=json \ --benchmark_out="$SCOPE_RESULTS/implicit_managed_HostWrGPU_coarse.json" -srun -c 56 -n 1 --gpus 8 $SCOPE_BUILD/comm_scope \ ---benchmark_repetitions=5 \ ---benchmark_filter='.*implicit_managed_GPUWrHost_fine/0/0/.*' \ ---benchmark_out_format=json \ ---benchmark_out="$SCOPE_RESULTS/implicit_managed_GPUWrHost_fine.json" - -srun -c 56 -n 1 --gpus 8 $SCOPE_BUILD/comm_scope \ ---benchmark_repetitions=5 \ ---benchmark_filter='.*implicit_managed_GPUWrHost_coarse/0/0/.*' \ ---benchmark_out_format=json \ ---benchmark_out="$SCOPE_RESULTS/implicit_managed_GPUWrHost_coarse.json" - -srun -c 56 -n 1 --gpus 8 $SCOPE_BUILD/comm_scope \ ---benchmark_repetitions=5 \ ---benchmark_filter='.*implicit_mapped_GPUWrHost/0/0/.*' \ ---benchmark_out_format=json \ ---benchmark_out="$SCOPE_RESULTS/implicit_mapped_GPUWrHost.json" - date diff --git a/scripts/run_12.sh b/scripts/run_12.sh new file mode 100755 index 0000000..a5f8ec8 --- /dev/null +++ b/scripts/run_12.sh @@ -0,0 +1,46 @@ +#!/bin/bash + +#SBATCH -A CSC465 +#SBATCH -J r12 +#SBATCH -o %x-%j.out +#SBATCH -t 2:00:00 +#SBATCH -p batch +#SBATCH -N 1 + +# salloc -A CSC465 -N 1 -p batch -t 2:00:00 + +export ROOT=/lustre/orion/csc465/scratch/cpearson/frontier-gpu-bandwidth +export SCOPE_SRC=${ROOT}/comm_scope +export SCOPE_BUILD=${ROOT}/build +export SCOPE_RESULTS=${ROOT}/run + +module load PrgEnv-amd/8.3.3 +export HSA_XNACK=1 + +mkdir -p $SCOPE_RESULTS +module list > $SCOPE_RESULTS/modules.r12.$SLURM_JOBID.txt 2>&1 +env > $SCOPE_RESULTS/env.r12.$SLURM_JOBID.txt +rocm-smi > $SCOPE_RESULTS/rocm-smi.r12.$SLURM_JOBID.txt 2>&1 +lscpu > $SCOPE_RESULTS/lscpu.r12.$SLURM_JOBID.txt 2>&1 + +date + +srun -c 56 -n 1 --gpus 8 $SCOPE_BUILD/comm_scope \ +--benchmark_repetitions=5 \ +--benchmark_filter='.*implicit_managed_GPUWrHost_fine/0/0/.*' \ +--benchmark_out_format=json \ +--benchmark_out="$SCOPE_RESULTS/implicit_managed_GPUWrHost_fine.json" + +srun -c 56 -n 1 --gpus 8 $SCOPE_BUILD/comm_scope \ +--benchmark_repetitions=5 \ +--benchmark_filter='.*implicit_managed_GPUWrHost_coarse/0/0/.*' \ +--benchmark_out_format=json \ +--benchmark_out="$SCOPE_RESULTS/implicit_managed_GPUWrHost_coarse.json" + +srun -c 56 -n 1 --gpus 8 $SCOPE_BUILD/comm_scope \ +--benchmark_repetitions=5 \ +--benchmark_filter='.*implicit_mapped_GPUWrHost/0/0/.*' \ +--benchmark_out_format=json \ +--benchmark_out="$SCOPE_RESULTS/implicit_mapped_GPUWrHost.json" + +date